Executive summary and key findings#
- Between 22 and 49% of B2B SaaS is invisible before retrieval: asked unprompted to name the leading tools in a category, GPT produced 78% of the cohort, Gemini 73, Claude Haiku 60 and an open Gemma 51. So the floor is set by an open model and the ceiling by the strongest closed one, and even there a fifth of real companies with real customers do not come up. Who is invisible barely changes between them, and on the open family recall climbed for three generations then stalled.
- One in five named brands is filed under the wrong category: of the 1k whose category we could verify, 194 are wrong. Accuracy then falls away rather than stopping there: of the 787 filed correctly, half are held below the cohort median margin, so the model prefers the right answer without holding it firmly. Confidently wrong is one verified fact in seventy, and it is the case a buyer cannot detect.
- Your own site buys accuracy, not recall: of the six channels we compared, your own properties predict being described correctly best, at 5.9 points per standard deviation, and being named almost least, at 1.4. A help centre alone adds 12 points of category accuracy, 3.8 once size is held constant. Documentation says what a product is and who it is for; a marketing page mostly asserts that it is good.
- Social audience buys 9.8 points of recall and 0.2 of accuracy: a large following gets you named and teaches the model nothing about what you are. Web prominence, crawl rank and referring domains, buys 12.6 points of recall and 7.9 of accuracy per standard deviation, so it is the only prominence that does both. Brands in the most prominent third of the open web are named 70% of the time against 40% for the least.
- Model sentiment has nothing to do with customer satisfaction: across 1k brands with G2 or Trustpilot listings, the correlation between real review scores and the sentiment in the weights is 0.03. The weights absorbed how a company is written about, not how well it serves its customers. This concerns parametric memory only: reviews still matter when a model retrieves and cites them at answer time, which is a different measurement with a different fix.
The 4 pillars of AI visibility measurement#
An AI answer is the end of a pipeline. The model reaches first for what its weights already hold, retrieval may or may not add documents on top, and only then is a response written. Almost every AI visibility tool built so far measures that final response and nothing else, which is the step closest to revenue but says nothing about how the model arrived there. We probe the whole system end to end instead, so that an absence comes with its cause rather than just its fact.
1. Parametric probing#
What the model carries in its own weights, before it has seen a single document. Fixed when training ends, identical for every user, and unchanged by anything you publish afterwards. Reading it means measuring the model rather than asking it, which needs open weights or the token probabilities most providers have now withdrawn.
It answers what a model already believes about you when nobody has handed it anything to read.
2. Mechanistic interpretability#
What the model does while the answer forms. The cheap read is chain-of-thought, taking the model at its word about where a brand fell out of the shortlist; the strong read is the computation itself, through logit lens and activation patching, since a model's account of its reasoning is a claim while its activations are evidence.
It answers where in the deliberation you were lost, which none of the others can.
3. Retrieval#
Not one fetch but a pipeline of its own. A question fans out into several queries, those are routed across different stores, and what comes back is filtered, deduplicated and trimmed to fit the context window. Being fetched is not being used: in one measured session the most-read source was never cited at all, and roughly two thirds of a page is cut before the model sees any of it.
It answers whether anything about you was in front of the model when it needed it, and if not, which step dropped it.
4. Response analysis#
What reaches the buyer, and what it cited. This is where most AEO tooling sits, because it needs no special access: ask repeatedly, record what comes back. It is also the most volatile of the three, since anything retrieval-backed changes the moment you publish.
It answers what a buyer sees today, and whether what they see is factually correct, deviated or simply unadded.
Where this report sits
Everything here is pillar one: what 1k public B2B SaaS brands look like inside model weights. It says nothing about how often you are cited in ChatGPT or AI Overviews, which is pillar three, has different rules and a different remedy, and which we measured separately across 2 million citations in What actually drives AI citations. They come apart in practice: a brand can be highly cited and nearly absent from the weights. For continuous measurement of the retrieval layer, that is our AI Visibility Tracker.
How models learn, and what they remember#
Training a frontier model runs in stages, from corpus assembly through pre-training, mid-training and post-training, and the stages matter to a brand differently, which is why we walk through them below.
Underneath every stage sits the same operation: the model reads the corpus a few tokens at a time, predicts the next token, and every wrong guess nudges billions of weights toward the right answer. Three variables set how much influence a token of your text ends up having on the final weights, and the stages above are the second of them:
- Repetition: how often the fact recurs across the corpus in independently phrased form. One page in a fifteen-trillion-token corpus, seen once, moves the weights about as much as one raindrop moves a river, while every independent restatement is another pass over the same weight directions, which is how the US Constitution's 218 training copies end up recitable. It is the strongest single dial, and the one we can verify against ground truth below.
- Training stage: when the passes happen. Passes do not simply accumulate, because what each one is worth depends on the learning rate at the time it lands: the same passage counts for more late in training, which is why labs anneal curated, high-quality corpora into the final stretch, and exposures seen only early partially fade as later training runs over them. Mid-training is the one stage where deliberate inclusion is possible.
- Pipeline processing: the data engineering between source and training run. Admission filters and deduplication decide whether text arrives at all, the channel that carried it (filtered crawl, licensed corpus, curated collection, synthetic rewriting) sets its odds, and reference-shaped content survives at far higher rates than conversational prose, which the source survey measures across 30 source types.
The reason to care is lifespan. A retrieved page is spent on one reply and competes for room while it lasts, where text imprinted in the weights takes part in every answer the model will ever give.
Probing the model for training-data signals#
Models are trained on a small slice of Common Crawl, not on the crawl itself. It is raw material, and a page has to survive corpus admission before any of it reaches the weights.
- Deduplication: exact copies collapse to one document, and near-duplicate detection prunes paraphrased clones alongside them.
- Perplexity filtering: text a small reference model finds bizarre is dropped, which catches both machine-generated filler and genuinely odd writing.
- Quality classifiers: a model trained on curated writing against random web text scores every document, so a good page formatted like a link farm can die here.
- And more: language identification, boilerplate and navigation stripping, toxicity and PII filters, and per-domain blocklists, each of which varies by lab and by run.
The published open pipelines show how aggressive the stack is. RefinedWeb discards roughly nine in ten raw documents before training, and the most-used subsets of FineWeb keep low single digits of a percent.
Two consequences matter to a brand, and the second is the one most content plans get wrong.
- The crawl is one source of several: licensed corpora, curated book and code collections and synthetic text never touch the public web pipeline, so they carry product discussion into training while diluting the open web marketers can actually reach.
- Syndication is a retrieval tactic, not a training one: a press release republished verbatim to forty sites survives admission as one document, so it can win citations at answer time while adding almost nothing to the weights. Training rewards independent restatement, the same fact reached and phrased differently by different authors on different domains.
The refinement does not stop at pretraining either, because training now runs in stages that differ in what they can do for a brand:
- Pre-training on the filtered web corpus is where facts about you enter the weights.
- Mid-training, which anneals curated technical corpora and long-context data into the final stretch of pre-training, is the one stage where a recognized authority in a narrow domain can be included deliberately rather than by crawl statistics.
- Post-training, the instruction tuning and preference optimization now built largely on expert-rated examples and model-generated synthetic data, adds no facts at all. It decides how what is already in the weights gets expressed, including whether a model names a specific brand or hedges into a category answer.
Probing the whole spectrum of models, open and closed#
Access falls away in steps. Open weights allow the full two-detector read, logprob APIs a partial one, and text-only output leaves just verbatim tail probing, near-conclusive when it fires and weak when quiet. This study runs across that whole spectrum.
The open models are not toy stand-ins. Gemma is distilled from the same pipeline as Gemini, and gpt-oss stands in the same relation to its closed siblings, so an open read is a floor on what the closed family knows. Distillation thins the tail, so a faint imprint in the teacher can vanish in the student, which is why every score is reported per model.
So asking what a model "knows" about a brand means asking how much nudging text about it caused, and how much of that trace survived the rest of training.
We call the depth of that trace the imprint. It is fixed by the training run and does not depend on who is asking, which is what makes it measurable rather than a matter of opinion.
What the crawl cannot tell you
Common Crawl is worth tracking, and Metehan's referring-domains web graph is a reasonable way to check it. But it covers one of several sources, and it tells you about the crawl rather than about the weights. Five questions sit downstream of it.
- Influence: how deeply your text marked the weights, not whether it appeared in a corpus.
- Accuracy: what the model asserts you are, and whether that is correct, deviated or unadded.
- Co-occurrence: which ideas, categories and companies the model holds next to your name.
- Sentiment: how it reads you on the dimensions a buyer weighs.
- Competitive: all four again for every rival, because a buyer picks one vendor from a shortlist.
ModelRank: a number for the imprint in the weights#
Most B2B SaaS companies have no read on their imprint at all. They know something is wrong when the model names a competitor instead, but not which of the 5 questions is the problem, and not what to do about any of them.
ModelRank
A 0 to 100, log-scaled reading of accumulated influence on a specific model's weights, taken per model and against the competitors you are chosen alongside. It is the carrier for the other 4 readings, and the number that says which of the 5 questions is your problem.
- Near 0: the model holds nothing it can be shown to have learned from you. At the floor the score reports uncertainty rather than pretending to a number.
- Mid range: your text left a trace the detectors can find but not one the model can reproduce. Nearly every brand in this study sits here.
- Near 100: the model can continue your text from a single line. This is where famous documents sit, and no B2B SaaS brand in the cohort reaches it.
The 4 readings under the score#
- Base imprint: how deeply the weights hold you. If training on your text nudged the model, the residue is readable, because the model is less surprised by your text than it has any right to be, especially at names and numbers that cannot be guessed from context. A second detector scrambles the surface form and watches whether confidence collapses, since understanding survives a paraphrase and memorisation does not. The two agree from different directions, which is the main reason to trust either.
- Co-occurrence: which categories, ideas and companies the model holds beside your name. This is the one you move directly, because it is built from what other people publish next to you, and it is the reason the other 2 read the way they do.
- Factual accuracy: whether the claims stored about you are yours. We ask the model to rank the true value against matched distractors, and the outcome is correct, deviated or unadded. When it deviates, it is almost always because a wrong category sits closer to your name than the right one, so the fix is co-occurrence, not a correction.
- Sentiment: the evaluative direction the training data taught the model to attach to you, read across 12 buying dimensions. Following Tigges et al. (2023), we fit that direction from contrast pairs and project your brand token onto it. A weak dimension is a dimension nobody writes about you on, which is again a co-occurrence problem.
Expand how the sentiment reading is taken
The direction is fitted once from contrast pairs, then every brand is projected onto the same axis. Each dimension uses its own prompt frame, so positions are comparable within a dimension and not across them.
Validating our system#
Every detector was tested on documents where membership is a checkable fact before being pointed at a brand. Against provably absent controls the raw detector scores 0.83, normalising against a reference model lifts it to 0.90, and on verbatim memorisation it separates perfectly at 1.00. A useless detector scores 0.50. The 0.90 is the operating number for everything in this report, and it means the detector ranks a true member above a true non-member 9 times in 10.
- Labelled sets: passages provably inside a public training corpus on one side, post-cutoff text and entity-swapped mutations on the other.
- Dose-response: run against Pythia, whose training data is published, with copy counts looked up exactly via infini-gram. The more copies the corpus held, the harder the detector fired: the US Constitution at 218 copies is reproduced verbatim by every provider we probed, while Moby Dick at 5.5 copies scores level with documents that were never there.
- Benchmark: the harness reproduces published results on the standard membership-inference benchmark.
- Prior work: memorisation and its growth with duplication are settled results from Carlini, Kandpal and Shi. What the literature does not do is point any of it at commercial brands.
Validation protocol and result files are available on request.
What is actually in the training data?#
With the instrument built, the question becomes a measurement rather than an argument. Of all the places a brand can publish or be discussed, which of them actually left a trace in the weights, and which are invisible no matter how much you put there? We sampled 2K+ passages from more than 30 kinds of source, from Stack Overflow answers to press releases, and scored each one for how strongly the model holds it.
Three angles on the same question
- Calibrating maximum imprint: books the model recites from a single line, which fixes the top of the scale so every reading below it is normalised against a known ceiling.
- Tracing what made it in: 30+ source types measured across 4 distilled model generations, using the training cutoff as a natural experiment to separate what was absorbed from what was merely crawled.
- The shape rule: what the surviving passages have in common, which turns out to matter more than which site they came from.
Given only the tail of a famous opening line, Gemma continues most of them word for word.
| Book | Given the opening line's tail, the model... |
|---|---|
| Pride and Prejudice (1813) | continues verbatim: "a good fortune, must be in want of a wife" |
| A Tale of Two Cities (1859) | continues verbatim through four clauses of the famous antithesis |
| Moby Dick (1851) | continues verbatim for 31 words |
| 1984, Harry Potter, The Hobbit, The Great Gatsby, The Hunger Games | each continues its opening word for word |
| Gone Girl (2012) | continues a mid-book passage word for word from a single line |
| The Da Vinci Code, The Martian | drifts into plausible but invented continuations: imprinted, not memorized at the surface |
That is what hundreds of independent copies buy, and it is the ceiling everything else in this report sits below.
Using model distillation to trace what made it into training data#
The training cutoff gives us a natural experiment. For each source we scored passages published before the cutoff, which the model could have trained on, against passages published after it, which it cannot have. A gap in favour of the earlier sample is evidence the source was ingested. No gap means the familiarity is style rather than memory.
The identical survey ran on 4 Gemma generations, all base pretrained models so post-training cannot reshape the read, which turns a snapshot into a time series of what the data pipeline kept.
These are distilled models, so a faint trace in the teacher can be compressed away in the student. A positive reading is evidence a source was ingested. A null reading is evidence the detector found no fingerprint, not proof the source was absent from training.
Key findings:
- YouTube leads every generation and is fading: +1.89, +1.61, +1.38, +1.18 across Gemma 1 to 4. Nothing has replaced it at the top; the whole open-web profile is flattening beneath it.
- What is rising instead is editorial and research: PubMed goes from no trace to +0.36, The Guardian from +0.32 to +0.75, Forbes from flat to +0.41. The pipeline is trading volume for provenance.
- Court opinions collapse hardest of anything measured: +1.43 to −1.61, a bigger reversal than LinkedIn's +1.53 to −0.58. Two of the deepest traces in Gemma 1 are gone by Gemma 4.
- Stack Overflow is the exception that tests the rule: the canonical reference-shaped source falls from +1.08 to zero on Gemma 4, while holding at +1.30 on Gemma 3 and returning at 31B. Being reference-shaped is necessary, not sufficient.
- One review platform survives and the rest do not: Amazon reviews hold at +1.05 while Trustpilot falls to −0.14 and G2 sits at zero. Reddit is null in every generation, including those trained before any licensing deal.
- The average fades but the variance is the story: mean era gap 0.551, 0.496, 0.442, 0.375 generation over generation, consistent with licensed and synthetic sources taking share, while individual sources swing far harder than that average suggests.
The shape rule: what survives the pipeline#
Pooling all 2K+ passages across every source, the strongest single predictor of survival is not which site the text came from but what shape it takes:
The pipeline behaves like a reference librarian with limited shelf space: it keeps what looks like something a future answer would cite, code, lists, and numbers, and passes over conversation. That preference operates whatever site the text came from, which is why a docs page on a small domain can outlive a thought-leadership post on a large one.
Takeaway
Write docs-shaped content: structured, numeric, specific, citable. The admission pipeline keeps reference material and discards conversation, so the same fact published as a spec table outlives it published as a post. And do not count on social platforms reaching the weights this generation; their influence runs through retrieval.
The audit: what AI knows about 1k B2B SaaS brands#
This section runs 1k public B2B SaaS brands, spanning 300+ categories from CRM to observability, through 2 baseline checks.
- Recall: does the model produce your name unprompted when asked what exists in your category.
- Explanation: is what it then says about you correct.
These are deliberately blunt, a yes or no on each. They tell you whether a problem exists, not what shape it is. The 4 readings under ModelRank, base imprint, co-occurrence, factual accuracy and sentiment, go deeper: they say how firmly the model holds you, what it holds beside you, how confidently it is wrong when it is wrong, and how it reads you against the competitors a buyer hears you beside.
5.1 Recall: do you exist when buyers ask?#
Per category, 2 unbranded prompts of the kind a buyer would use ("the most widely used CRM software tools for businesses are") and 6 completions each. A brand is recalled if its name appears anywhere in those 12.
Key findings:
- Waiting for the next model does not work: unprompted recall climbs from 39.8% on Gemma 1 to 44.3 on Gemma 2 and 51.1 on Gemma 3, then stops: Gemma 4 lands at 51.0, a dead heat with its predecessor at the same size on the identical cohort. Three generations of gains came free and the fourth did not, so the 490 brands that go unnamed should be treated as a problem to work on rather than one the next release will absorb.
- It is also worth checking per assistant: the same cohort is named unprompted 51% of the time by Gemma 4, 60% by Claude Haiku 4.5, 73% by Gemini 3.5 Flash Lite and 79% by GPT-5.6 Luna, so the invisible half shrinks to roughly a fifth on the strongest model.
- Your category caps your odds, so measure against it and not the cohort: project management surfaces 8 of its 12 brands and CRM 7 of 12, while education tech and databases surface 2 of 10 or 11. A crowded category is not a visible one, and the incumbents' grip on the answer space is the number your recall should be benchmarked against.
- One check is not a reading: reseeding the same prompts flips 13% of individual brands between named and not named. A single screenshot of an assistant naming you, or failing to, is inside the noise. Sample the category phrasings your buyers actually use, repeatedly, and track the trend rather than the instance.
Takeaway
The question that matters is not "does AI know my brand" but "does AI name my brand when a buyer asks about my category". Track it the way you would track rankings: fix the category phrasings your buyers use, run unprompted recall against them quarterly, and treat the incumbents' grip on your category's answer space as the competitive baseline to erode through independent restatement.
5.2 Facts: what it believes about you#
For every brand we take a fact we can verify independently, its product category, and make the model rank the true value against 4 distractor values of the same type and length. Across the cohort 787 brands are placed correctly, 194 are placed in the wrong category and 19 hold nothing either way, so for 1 in 5 real software companies the model does not merely lack an opinion about what they sell, it holds the wrong one.
- Consistent: the true value wins, and the margin over the runner-up says how firmly.
- Unaddressed: nothing wins by a meaningful margin, which is materially different from being wrong.
- Deviated: a wrong value wins, usually because text about a different company was absorbed under your name.
- The dangerous state is common and invisible: only 14 facts in 1k are confidently wrong, but another 180 are wrong without signalling any doubt to the reader. Nothing in an assistant's answer tells a buyer which band it came from, so a model sounding certain about you is no evidence that it is right.
- Being famous does not protect you: the model files Okta under CI/CD rather than identity, Square under card issuing rather than SMB commerce, Akamai under CI/CD rather than CDN and HashiCorp under workflow orchestration rather than infrastructure security. Check your own entry rather than assuming footprint has done the work.
- A correct answer can still be one campaign from flipping: 84 of the 787 correct placements win by less than 0.2, with Sumo Logic holding observability by 0.020 and Conductor holding SEO by 0.024. Correct is a position to defend, not a box to tick.
- Whether an assistant commits or declines is a property of the model, not of you: we ran the same 1k brands through GPT-5.6 Luna, Claude Haiku 4.5 and Gemini 3.5 Flash Lite as well. As a rate of being wrong when it commits, Gemma 4 sits at 20% against 2% for Claude, 0.8 for Gemini and 0.4 for GPT, because GPT declines on 287 brands while Gemini declines on 8. Test your brand on each assistant separately: one that declines to place you needs presence, one that misfiles you needs correction, and the same content programme will not fix both at the same rate.
- Read that spread as flattering the closed models: Gemma ranks 5 sentences with no options in front of it and no way to abstain, while the closed models see the options and are offered an explicit "I do not know", so the accuracy figures are not like for like.
- A quarter of the cohort carries nearly all of the error: 742 of 1k brands sit in confidence bands where the model is right 90% of the time and 609 where it is right 95%, so the remaining 258 hold almost every wrong answer. Either your brand is in that tail, where one check will find it, or the model's read on you is already stable. A wrong or weakly held answer covers over 400 brands in all, and for them the gap is not competitive: nobody currently occupies the correct answer, so the first clear, independently restated statement of what the company does has very little to displace.
- Models know what you are and who you connect to, and almost nothing about how you sell: product category is right 80% of the time against a 1 in 5 chance baseline, and integrations separate true from false pairs at 0.61 against a coin's 0.50. Free tier lands at 0.55, published pricing at 0.53 and self-serve at 0.45, which is worse than chance. Pricing is among the most common things a buyer asks an assistant, and there is nothing in the weights to answer it with.
- Restatement by someone else is what makes a fact stick: an integration gets written down by both parties, quoted in comparisons and listed in changelogs, while pricing appears once, on your own page. Your own site moves being filed correctly by 8.9 points and a help centre adds 12, but it moves recall by under 3 with the interval straddling zero. So publishing a commercial fact is what makes it legible, and getting it restated somewhere you do not own, starting with the partner's own directory listing, is what makes it known.
- Wrong answers land on a handful of magnet categories: the 194 deviations hit only 94 categories, with 10 absorbing a third of them. CI/CD attracts 9 misfiled brands while holding 2 of its own, and web hosting attracts 6 while holding 1. Compact technical labels the training text uses constantly act as attractors for brands the model cannot place.
- Misfiling is a vertical problem, and it ranks: the share of a group's brands filed under the wrong category, worst first.
- Regulated verticals, health, education and government: 37%.
- Developer tools: 30%.
- Customer experience, and finance and legal: 24% each.
- Sales and marketing: 18%, with HR and operations at 16%.
- Security and identity: 12%, with data and analytics and product and design at 10%.
- IT and internal ops: 0%, every brand in the group filed correctly.
5.3 Sentiment: who reads well, who reads badly, and on what#
Model sentiment is not customer satisfaction: star ratings and review counts from G2 and Trustpilot were available for 1k of these brands, and the correlation between a brand's real customer rating and the sentiment held in the weights is 0.03, which is nothing. Review sites are a citation-time lever rather than a training-time one: an assistant retrieves and quotes them while answering, which is why they move what a buyer sees today and not what the weights hold. What bakes sentiment in is prose about the company, and the training-data results say which surfaces carry it.
- Press writes in the register sentiment is made of: a TechCrunch snippet states what a brand actually does 33.9% of the time and a Forbes snippet 27.8%, so coverage there carries a characterisation and not just a name.
- Developer surfaces name you without describing you: GitHub states the category 14.9% of the time and Stack Overflow 9.3%. A dependency manifest records that a tool exists, not what it is like to live with, which is why these surfaces move recall far more than they move how you read.
- Authority decides whether the characterisation survives: Common Crawl graph rank, referring domains and Wikipedia citations of your domain are the 3 variables that move both being named and being described correctly, so a flattering account of you needs to sit somewhere with the authority to be restated.
Every ranking below is therefore a record of how a company is written about, not how well it serves the people paying for it.
- Adoption experience drives it, not price: onboarding separates the best-regarded hundred brands from the worst by 3.3 standard deviations, ease of use by 3.2, reliability by 3.1 and documentation by 3.0. Pricing separates them by 1.8, the narrowest of the twelve. Whatever the model absorbed about these companies is weighted toward how a product feels to adopt and run.
- One judgement carries most of it, but not all: a single factor accounts for 55% of the variance and the median correlation between any 2 dimensions is 0.52, so a general reputation moves much of the set at once.
- Some pairs are a single lever: documentation and onboarding move together at 0.89, so there is no realistic route to shifting one without shifting the other.
- The rest is still worth reading separately: at a median of 0.52 any 2 dimensions share only about a quarter of their variance, so the per-dimension readings carry real information about where you are weak. They just will not move fully independently of the headline.
Who leads and who lags, category by category
Two patterns run through the tables below. The brand that reads best is frequently not the market leader: Zoho above the rest of CRM, Moz above Ahrefs in SEO, Monday.com above Atlassian in project management. And the laggard's weakness is usually an experience dimension, support, onboarding or ease of use, which is consistent with the driver chart above. Pick a category to see every brand in it, ranked by how the model reads them.
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Insider | +1.26 | product depth | innovation | |||
| Customer.io | +0.71 | ease of use | integrations | |||
| OneSignal | +0.66 | community | support | |||
| Iterable | +0.65 | pricing | reliability | |||
| Klaviyo | +0.52 | ease of use | reliability | |||
| CleverTap | +0.46 | community | support | |||
| Braze | +0.38 | community | support | |||
| Postscript | +0.30 | onboarding | pricing | |||
| ActiveCampaign | +0.23 | ease of use | trajectory | |||
| Attentive | +0.10 | onboarding | community | |||
| Marketo | -0.17 | support | onboarding | |||
| Dotdigital | -0.93 | support | community | |||
| MoEngage | -0.95 | pricing | integrations | |||
| Mailchimp | -0.96 | product depth | support | |||
| Bloomreach | -1.55 | trajectory | community |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Moz | +1.42 | product depth | innovation | |||
| Clearscope | +1.05 | community | pricing value | |||
| Botify | +0.88 | community | trajectory | |||
| Discovered Labs | +0.79 | ease of use | trajectory | |||
| Peec AI | +0.62 | reliability | integrations | |||
| Similarweb | +0.54 | support | pricing value | |||
| Conductor | +0.46 | reliability | pricing value | |||
| Lumar | +0.41 | support | documentation | |||
| Surfer | +0.26 | ease of use | trajectory | |||
| SE Ranking | +0.26 | support | pricing value | |||
| Frase | +0.17 | support | trajectory | |||
| MarketMuse | +0.09 | trust and security | pricing value | |||
| Semrush | +0.03 | community | pricing value | |||
| BrightEdge | -0.09 | support | integrations | |||
| Ahrefs | -0.20 | community | ease of use | |||
| Profound | -1.93 | product depth | onboarding |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Zoho | +1.49 | product depth | integrations | |||
| HubSpot | +1.14 | product depth | innovation | |||
| Attio | +0.95 | reliability | trust and security | |||
| Close | +0.52 | pricing | reliability | |||
| SugarCRM | +0.29 | support | pricing | |||
| Creatio | +0.29 | onboarding | pricing | |||
| Insightly | +0.12 | documentation | trust and security | |||
| Salesforce | +0.07 | trajectory | product depth | |||
| Freshworks | +0.00 | support | community | |||
| Copper | -0.09 | trajectory | support | |||
| Keap | -0.49 | support | community | |||
| Pipedrive | -0.59 | product depth | support |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Tidio | +1.43 | community | pricing | |||
| LiveChat | +1.30 | ease of use | integrations | |||
| Gladly | +1.09 | community | support | |||
| Kustomer | +0.98 | onboarding | support | |||
| Gorgias | +0.94 | community | trajectory | |||
| Freshdesk | +0.58 | ease of use | pricing | |||
| Intercom | +0.55 | documentation | trajectory | |||
| Front | +0.47 | pricing | support | |||
| Zendesk | +0.32 | documentation | support | |||
| Help Scout | +0.27 | innovation | integrations | |||
| Hiver | -0.17 | ease of use | trust and security | |||
| Pylon | -0.48 | trust and security | support |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Monday.com | +1.43 | reliability | integrations | |||
| ClickUp | +0.66 | community | pricing | |||
| Asana | +0.62 | community | pricing | |||
| Basecamp | +0.51 | innovation | product depth | |||
| Linear | +0.50 | trajectory | trust and security | |||
| Trello | +0.48 | community | pricing | |||
| Shortcut | +0.17 | trajectory | support | |||
| Wrike | +0.05 | community | innovation | |||
| Height | -0.03 | documentation | support | |||
| Teamwork | -0.19 | reliability | community | |||
| Smartsheet | -0.32 | product depth | trajectory | |||
| Atlassian | -1.46 | pricing | onboarding |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Paddle | +1.44 | product depth | innovation | |||
| Adyen | +0.79 | community | trajectory | |||
| Flywire | +0.64 | pricing | trajectory | |||
| Stripe | +0.60 | pricing | reliability | |||
| Checkout.com | +0.35 | onboarding | trajectory | |||
| Wise Platform | -0.01 | trust and security | documentation | |||
| Payoneer | -0.44 | documentation | trajectory | |||
| Rapyd | -0.50 | support | ease of use | |||
| Lemon Squeezy | -0.79 | pricing | integrations | |||
| GoCardless | -0.81 | pricing | integrations | |||
| Airwallex | -1.00 | product depth | ease of use |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Make | +0.81 | product depth | support | |||
| Tray.ai | +0.76 | documentation | integrations | |||
| Paragon | +0.66 | pricing | reliability | |||
| Merge | +0.52 | pricing | reliability | |||
| Zapier | +0.34 | onboarding | reliability | |||
| Workato | +0.06 | support | product depth | |||
| n8n | -0.39 | pricing | integrations | |||
| SnapLogic | -0.40 | support | innovation | |||
| MuleSoft | -0.55 | support | innovation | |||
| Boomi | -0.57 | support | innovation | |||
| Celigo | -1.14 | support | community |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Tableau | +1.15 | documentation | onboarding | |||
| Zoho Analytics | +0.62 | onboarding | pricing | |||
| Metabase | +0.61 | innovation | trajectory | |||
| Domo | +0.39 | reliability | innovation | |||
| Sigma Computing | +0.37 | onboarding | trajectory | |||
| Hex | +0.24 | innovation | support | |||
| Mode | +0.23 | pricing | trust and security | |||
| ThoughtSpot | -0.35 | support | documentation | |||
| Qlik | -0.45 | trust and security | integrations | |||
| Looker | -0.46 | pricing | reliability | |||
| Sisense | -0.53 | support | documentation |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Anthology | +1.26 | product depth | innovation | |||
| Kahoot! | +0.70 | community | pricing | |||
| Panorama Education | +0.45 | onboarding | trajectory | |||
| Clever | +0.40 | pricing | trust and security | |||
| Turnitin | +0.13 | onboarding | trajectory | |||
| Nearpod | -0.20 | trust and security | onboarding | |||
| PowerSchool | -0.45 | onboarding | support | |||
| D2L | -0.53 | pricing | community | |||
| Instructure | -0.86 | pricing | community | |||
| Securly | -0.99 | trust and security | community | |||
| GoGuardian | -1.55 | support | community |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Better Stack | +0.86 | onboarding | integrations | |||
| AppDynamics | +0.01 | support | reliability | |||
| Chronosphere | -0.21 | pricing | documentation | |||
| Honeycomb | -0.29 | ease of use | trajectory | |||
| Grafana Labs | -0.30 | community | integrations | |||
| Sumo Logic | -0.35 | support | reliability | |||
| Datadog | -0.74 | product depth | integrations | |||
| New Relic | -0.85 | product depth | ease of use | |||
| Dynatrace | -0.92 | product depth | integrations | |||
| Splunk | -0.95 | community | ease of use |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| Neon | +1.37 | onboarding | community | |||
| Supabase | +0.36 | ease of use | support | |||
| MongoDB | +0.28 | pricing | trust and security | |||
| Timescale | +0.23 | support | documentation | |||
| PlanetScale | -0.07 | support | documentation | |||
| ClickHouse | -0.28 | pricing | documentation | |||
| Redis | -0.30 | documentation | support | |||
| InfluxData | -0.41 | onboarding | community | |||
| CockroachDB | -0.52 | product depth | integrations | |||
| SingleStore | -0.85 | support | documentation |
| Brand | Sentiment | Delta vs cohort | Distinctive for | Weakest on | Recall | Category fact |
|---|---|---|---|---|---|---|
| SentinelOne | -0.21 | support | innovation | |||
| Bitdefender | -0.42 | support | ease of use | |||
| Check Point | -0.49 | onboarding | trajectory | |||
| Palo Alto Networks | -0.59 | support | integrations | |||
| Zscaler | -0.60 | trajectory | ease of use | |||
| Malwarebytes | -0.94 | trust and security | community | |||
| CrowdStrike | -1.03 | support | innovation | |||
| Sophos | -1.07 | trajectory | community | |||
| Trellix | -1.09 | pricing | integrations | |||
| Netskope | -1.40 | product depth | innovation |
Brand knowledge: where does it come from?#
The audit says what the model holds. This section asks what put it there, running approximately 80 variables against the same 2 outcomes to see which of them predict recall and which predict being described correctly.
Initial observations, variable by variable#
Running the same approximately 80 variables against both outcomes, each in its own regression holding crawl prominence and referring domains constant, separates them cleanly. Testing only explanation was why every coverage measure had looked dead.
- An ordinary-word name is the strongest effect and it is a liability: it costs 14.1 points of recall and 32.0 points of accuracy, and brands called Close, Chorus, Clay or Terminus are filed correctly 56.3% of the time against 88.2% for everyone else. The model has to hold your brand in the same representational space as the ordinary word and the word usually wins, putting Close in CI/CD, Chorus in e-commerce marketing, Clay in data observability and Terminus in CI/CD. It is also the one variable here a company cannot change after the fact.
- A crowded category costs recall, not understanding: the sparsest third of categories are recalled more often than the most crowded, while the crowding effect on being filed correctly disappears entirely, at a 1.5-point gap indistinguishable from noise. More names competing for the same slot means less answer space for yours, not a muddier idea of what you do.
- Your own publishing buys accuracy, not existence: how much of your own site the model absorbed moves recall by under 3 points with the interval straddling zero, but moves being filed correctly by 8.9 points, and having a help centre adds 12 on the same outcome because a knowledge base states plainly what a product does. Your own publishing is how the model learns what your words mean, not how it learns that you exist.
Grouping the variables into composites#
Testing approximately 80 variables one at a time makes a handful look significant by chance and hides the pattern underneath, so we grouped them into 8 composites, each shaped like a channel a marketing team actually owns, with the individual variables plotted below each one so nothing is hidden inside an average.
2 of the 8 measure prominence itself, and every other composite is adjusted for them, which is a demanding test that removes about half of what looks significant alone.
- Common Crawl prominence: rank in the open-web link graph, a proxy for how much of the web points at you at all. Worth 13 points of recall and 8 of accuracy per standard deviation, the largest effect measured on either outcome.
- Social prominence: audience size across the 8 platforms we measured. Worth 10 points of recall and nothing at all on accuracy, which is the clearest single case of reach without comprehension.
The other 6 are the channels a team can actually spend against.
- Social presence: followers, audience and co-occurrence across Reddit, YouTube, LinkedIn, X, Instagram and Facebook.
- Technical authority: mentions, stars and co-occurrence across GitHub, Stack Overflow, Hacker News, dev.to and Medium.
- Co-occurrence coverage: the share of all pages mentioning a brand that also state its category.
- Review sites: mentions, co-occurrence and review volume on G2 and Trustpilot.
- News coverage: mentions and co-occurrence across the New York Times, the Guardian, Forbes and TechCrunch.
- Your own properties: help centre and its depth, documentation, pricing disclosure, integrations directory, how familiar the model finds your own site, and presence in Common Crawl.
The result splits by outcome. Technical authority is what gets a brand named, at 5.4 points per standard deviation, with co-occurrence coverage, review sites and news coverage at the boundary and social presence doing nothing. Your own properties are what get a brand understood, at 5.9 points on category accuracy, the strongest single result on that outcome, while technical authority runs slightly negative there, which is the GitHub pattern again: code mentions carry a name without ever saying what the product is.
Social presence doing nothing on either outcome is worth sitting with, because the engines appear to compensate for it at the other end of the pipeline. Reddit leaves no fingerprint in any Gemma generation, yet it takes 27% of ChatGPT's search slots during query processing, which we measured separately in Reddit and LLM citations. It reads like a separation of concerns: what the weights hold and what gets retrieved at answer time are doing different jobs, and a source absent from one can dominate the other.
What predicts being named#
- Reach dominates recall, and it is audience rather than ownership: follower counts on Facebook, YouTube and LinkedIn each move recall by roughly 30 points, as do referring domains and rank in the Common Crawl graph.
- Simply having a channel does nothing, which is why the measures of what a brand maintains came back empty.
- A large audience produces repeated third-party mention across many domains, which is what ingestion rewards, so followers are a proxy for that volume rather than a cause.
- A brand can be prominent inside one platform or region and still leave little trace in a corpus weighted toward what is linked and quoted more widely.
- Co-occurrence evidence: brands whose mentioning pages also state their category are recalled 56.6% of the time against 47.0%. Adjusted, that is roughly 4 points of recall per standard deviation, which misses our correction at 0.077 but clears it outright on Stack Overflow and dev.to.
- How many times you are mentioned beats whether you are mentioned: counting Hacker News mentions moves recall from 40.8 to 58.8%, and Stack Overflow from 41.6 to 59.3%. Both are stronger than asking the same question as a yes or no, so being present on a surface matters far less than how often.
Outcome one, every variable tested. Hollow markers did not survive correction within their family.
What predicts being described correctly#
- What the coverage says matters more than how much of it there is: the total number of pages mentioning a brand does nothing on accuracy, the share of those pages titled after the brand costs 10.7 points, and long, rambling mention text costs 22.7.
- Coverage that names your category tracks with being filed correctly: those brands are right 86.2% of the time against 72.4%, a gap of 13.8 points, though prominent brands have both so this is an association rather than a demonstrated cause.
- Co-occurrence only pays on a few surfaces: of the 17 we tested, coverage that names your category alongside your brand moves an outcome on 5. Everywhere else it does nothing.
- Gets you named: Stack Overflow at 18.7 points, then dev.to at 16.7.
- Gets you described correctly: Reddit at 13.1, then Forbes at 11.6, then YouTube at 9.6.
Outcome two, every variable tested.
- A Wikipedia article gets you named and stops you being understood: having one moves recall up 10.4 points and accuracy down 16.7, and a longer article carries the same double edge, up 13.0 and down 14.7.
- An encyclopaedia entry describes a company historically and corporately, in funding rounds and acquisitions, which teaches a model your name without teaching it what you sell.
- It also ages badly. Entries are often years behind a company's current product and positioning, so the model learns a version of you that stopped being true, and it learns it firmly.
- The fix is simple and cheap: keep the article current, and make sure the opening line says what the product is and who it is for rather than when the company was founded.
- Authority is the one prominence that buys both outcomes: rank in the Common Crawl graph moves recall by 29.5 points, referring domains by 29.8 and Wikipedia citations of your domain by 24.4, all overwhelming, and against accuracy the same 3 move it by 18.7, 17.2 and 13.2 and now survive correction. What separates the outcomes is therefore which of the remaining variables survive once authority is held constant. Both charts plot every variable tested, so a null is visible as a hollow marker rather than simply missing.
- Developer surfaces name brands without explaining them: a TechCrunch snippet names the brand's category 33.9% of the time and a Forbes snippet 27.8, against 14.9% for GitHub and 9.3 for Stack Overflow. A dependency manifest records that a tool exists without saying what it is for, which is why developer surfaces move recall while doing much less for accuracy.
How to get your brand into the training data#
The pillars fail independently, so you can be cited constantly and nearly absent from the weights, or deeply imprinted and described wrongly. They compound when they hold together, though, which is what separates the brands that dominate a category from the ones that merely appear in it: an imprint that is deep, correct and favourably described makes every citation land harder, and the citations feed the next round of training. The work therefore starts with a diagnosis of which pillar is missing rather than with a tactic.
Diagnose which problem you have#
2 ways to take the reading, in descending order of rigour.
- A ModelRank read: recall, category accuracy and the 12 sentiment dimensions, measured in the weights against your competitive set. Our free ModelRank audit, which runs it on your own brand, is coming soon.
- Paired with the answer layer: the weights are only pillar 1, so response and chain-of-thought analysis belong alongside, which is what our AI Visibility Tracker runs continuously across ChatGPT, Claude, Perplexity and Google AI Overviews.
Whichever route you take, the reading puts you in one of 4 rows, and each row is a different problem with a different fix.
| Symptom | Layer | Fix |
|---|---|---|
| Never named unprompted | Imprint | A bet on the next training cycle: independent, reference-shaped coverage on the technical surfaces that survive ingestion, with your category stated next to your name. |
| Named, but filed under the wrong category | Binding | Fix your own docs, help centre and pricing pages, the strongest accuracy channel we measured, and get disambiguating third-party coverage that pairs name with category. Doubly urgent if your name is an ordinary word. |
| Named and filed correctly, but read poorly | Experience | Third-party writing about onboarding, reliability and documentation, the dimensions where the model actually discriminates. |
| Strong in the weights, weak in AI answers | Retrieval | A different problem with different rules, out of scope here: our AI Visibility Tracker measures that layer continuously. |
Opportunities to grab early#
- The default answer in most categories is still unoccupied: 490 of 1k brands are never named, and a wrong or weakly held category answer covers over 400 more. For most of this cohort the first clear, independently restated statement of what a company does has almost nothing to displace, which makes now the cheapest this will ever be. Both counts come from the audit.
- Pricing is uncontested because nothing holds it: the weights separate true from false published-pricing claims at 0.53, barely better than a coin, so every AI answer about your pricing is produced from retrieval rather than memory. A crawlable, plainly worded pricing page wins that answer outright, and no competitor can out-imprint you on it.
- A help centre is the highest-return page you probably do not have: having one adds 12 points of category accuracy on its own and integrations pages add 4.6 adjusted, both fully in your control and neither needing anyone else's permission, and both measured in where brand knowledge comes from.
Weaknesses to fill in#
- Check whether you are misfiled before anything else: 1 in 5 brands are filed under the wrong category outright, rising to 37% in regulated verticals and 30% in developer tools. Being wrong is a different problem from being absent, and no amount of reach fixes it. The full ranking by vertical is in the audit.
- Fix it with coverage that states what you do, not more coverage: the share of pages that name you and also say what you do is the one content property with independent signal on recall, worth roughly 4 points per standard deviation and sitting at the boundary of significance under our strictest correction. Track it like a KPI anyway, because it is the disambiguation mechanism where misfiled and ordinary-word brands recover, and raw mention volume does nothing for accuracy. The channel-by-channel evidence is in where brand knowledge comes from.
- Do not expect your own site to solve existence: how much of it the model absorbed moves being described correctly by 8.9 points but moves recall by under 3, with the interval straddling zero. Documentation fixes what your words mean, not whether you are there.
Strengths to back up#
- Technical authority is the only channel that clears every test on being named: 5.4 points per standard deviation with company size held constant, across Stack Overflow, dev.to, Hacker News and GitHub-adjacent surfaces. If you already have developer traction, this is the lever with the least resistance in front of it.
- Authority is the only prominence that buys both outcomes: rank in the Common Crawl graph moves recall by 29.5 points and being described correctly by 18.7, referring domains 29.8 and 17.2, and Wikipedia citations of your domain 24.4 and 13.2. Every other lever pays into one outcome and not the other, as where brand knowledge comes from sets out in full.
- A correct answer is a position to defend, not a box ticked: 84 of the 787 correctly placed brands win by less than 0.2, with Sumo Logic holding observability by 0.020 and Conductor holding SEO by 0.024. If that is you, the spend is maintenance rather than acquisition.
Competitive threats#
- Thin margins flip both ways: the 0.2 margin that makes your position defensible makes a competitor's winnable, and 103 of the 194 wrong answers in the cohort still clear it, so half of them reach a buyer as a specific claim rather than a hedge.
- Your category caps you and the incumbents set the cap: project management surfaces 8 of its 12 brands while education tech surfaces 2 of 11. Benchmark against your own category's rate rather than the cohort's, and treat the incumbents' grip on the answer space as the thing to erode.
- Waiting is the losing move: recall climbed 11 points across 3 Gemma generations and then stalled at 51.0, which the audit shows generation by generation. The imprint compounds for whoever is publishing now, and the next release will not hand it to you.
Keep score#
Measure quarterly, per model, against your named competitor set. Quarterly is the right cadence because it matches the layer this report measures, and the layers of a model move at very different speeds.
- Alignment, weekly: RLHF and instruction tuning are updated more or less continuously, which is why an assistant's tone and its willingness to name anyone at all can change between one week and the next without anything having changed about you.
- Continued pretraining, monthly to quarterly: this is the layer where new text actually reaches the weights, so it is the one your coverage is buying into and the one worth reading on a quarterly cycle.
- A genuinely new base model, every 3 to 6 months at present: the step changes in recall between generations land here, and they are the payoff for coverage published well before the training cut-off.
So this is a compounding play rather than a week-by-week one. Nothing you publish this week shows up in the weights this week, and the work that does land keeps paying across every subsequent generation. The most repeatable finding in this report is that the default answer in most categories is still unoccupied. This is the programme we run for clients, most visibly for Gladia, whose sales-accepted leads from AI search grew sevenfold, and ModelRank, the instrument we built to run it, is coming soon.