AI brand recall: what LLMs know about your brand before they search

TL;DR:

• Discovered Labs' study of 1,000 B2B SaaS brands found two separate failure modes in AI brand recall: models that never name a brand, and models that name it but file it under the wrong category.
• Recall from memory alone ranged from 51% of brands on Gemma 4 to 78.5% on GPT-5.6 Luna, so between a fifth and half of the B2B SaaS brands tested are missing from model memory.
• Missing from memory: earn consistent third-party mentions. Expect this to build over several model releases.
• Named but misclassified: align category language across your site, G2 profile, and comparison content. You're correcting an existing association, so it often moves faster.
• Diagnose first: run direct ("What is [Brand]?") and reverse ("best [category] tools") queries with web search off, then fix the broken layer.

B2B buyers increasingly research vendors inside AI assistants, and the answer they get depends partly on what the model already knows about your brand. Some brands are missing from that memory entirely, while others are known but filed under the wrong category. This guide is for marketing leaders who own organic search. It covers how AI brand recall works, how to measure it, and which fix to make first.

AI brand recall is what a large language model knows about your brand from its training data, before it searches the web. It has two parts: whether the model names your brand at all, and whether it places you in the right category. In Discovered Labs' study of 1,000 B2B SaaS brands, Gemma 4 named 51% of brands from memory and misfiled 19.8% of its category choices. Recall and category accuracy fail separately, so each needs its own fix.

How AI brand recall shapes buyer research#

AI brand recall decides whether your brand can appear in an answer before the assistant looks anything up. That matters because buyers now do much of their vendor research in AI assistants. In G2's 2026 survey of 1,076 B2B software buyers and decision-makers, 71% rely on AI chatbots at some point in their research.

How parametric memory differs from retrieval#

Parametric memory is information encoded in a model's weights during training. The model regenerates facts from patterns in its parameters, much like answering a closed-book exam.

Retrieval-augmented knowledge arrives at query time. The model searches an external source, such as a web search index or a vector index of documents, and synthesizes an answer from the passages it finds.

Retrieval-augmented generation pairs the knowledge models store in their parameters with this external memory. The original RAG paper names updating a model's stored world knowledge as an open research problem.

In practice, the two layers move at different speeds. Parametric memory changes only when a lab trains a new model, while retrieval responds to content the model can find today. When you ask a model "What is Salesforce?", it regenerates facts from learned links between Salesforce, CRM, and enterprise software. Retrieval adds to that core, but the two layers don't always agree.

How buyers decide inside AI answers#

Many searches never send buyers to a website at all. In Pew Research Center's browsing data, Google users clicked a traditional result in 15% of visits without an AI summary, and in only 8% when one appeared.

Buyers also act on what the assistant recommends. In the same G2 survey, 69% chose a different vendor than they had planned because a chatbot recommended it.

This can undermine traditional attribution. If an assistant recommends a competitor because your brand isn't in its memory, that buyer may never reach your site. In these cases, the consideration set can form inside the AI assistant.

What a 1,000-brand study shows about LLM memory#

Discovered Labs' brand recall study found that recall and category accuracy are separate failure modes. Recall asks whether a model names your brand at all. Category accuracy asks whether it places you in the right category.

How the study measured brand knowledge#

Discovered Labs tested 1,000 B2B SaaS brands on Gemma 4, with cross-checks on GPT-5.6 Luna, Gemini 3.5 Flash Lite, and Claude Haiku 4.5, to measure what models know before web search. The brands span 333 product categories.

Recall was tested with unbranded category prompts, such as "the most widely used CRM software tools for businesses are," counting whether the model named the brand. Category accuracy asked the model to pick the brand's verified category from five candidate categories.

Each prompt was sampled several times, 12 completions per category on Gemma 4, because answers vary between runs.

Which brands models remember#

Recall depended heavily on the model. Gemma 4 named 51% of the brands (510 of 1,000) from memory alone. The closed models named more: 78.5% for GPT-5.6 Luna, 73.3% for Gemini 3.5 Flash Lite, and 60.2% for Claude Haiku 4.5.

Brands that achieved recall tended to have wider reach: more social followers, more Hacker News and Stack Overflow mentions, and more prominence in Common Crawl. Brands with ordinary-word names were recalled less often.

Where models misclassify brands#

On Gemma 4, 19.8% of category choices (194 of 981) differed from the brand's verified category. The closed models misfiled far fewer, about 2% for Claude Haiku 4.5 and under 1% for Gemini and GPT. They could decline to answer while Gemma 4 could not, so the rates aren't directly comparable.

In our experience, misclassification can do more damage than absence. A model that doesn't know you costs you a discovery opportunity. A model that misfiles you can send buyer intent toward competitors in an adjacent category.

Direction matters too. LLMs trained on a fact in one direction often fail to learn the reverse, a pattern researchers call the reversal curse. A direct query ("What features are in Brand X's enterprise tier?") can surface the right association. A reverse discovery query ("What are the best enterprise tools for [Feature Y]?") may default to the best-known names in the category.

How to decide which layer to fix first#

Diagnose which layer is broken before you invest in content or technical work. Recall and category accuracy have different symptoms, causes, and fixes.

Diagnose which layer is broken#

Work through these four checks in order:

1. Run a baseline audit. Query ChatGPT, Claude, and Gemini with direct and reverse queries, with web search turned off or through the API with no tools. Treat any answer that carries source citations as retrieval. Perplexity is built around live web search, so use it to test retrieval. Track whether your brand is named and whether the category assignment is correct.
2. Check your knowledge cutoff status. If your brand launched or repositioned after the model's cutoff, you're in retrieval territory. Focus on content and third-party mentions while you wait for the next model release.
3. Audit third-party consistency. Search for your brand on Reddit, G2, industry publications, and comparison sites. Note each category label used to describe you.
4. Compare retrieval, recall, and category. If you're cited in retrieval-augmented answers but not named from memory, you have a training data problem. If you're named but misclassified, you have a consistency problem.

Fix missing recall#

Symptom: The model doesn't name your brand in direct or reverse queries.

Cause: Too little presence in training data. Your brand launched after the model's knowledge cutoff, or it lacks consistent third-party mentions.

Fix: Build third-party consistency. Earn mentions in industry publications, comparison content, and community discussion. Brand mentions earned between two model versions' cutoffs can enter the newer model's memory.

Timeline: Expect parametric change to build over several model releases.

Fix misclassification#

Symptom: The model names your brand but files you in the wrong category.

Cause: In our experience, the usual cause is inconsistent category language across your content and third-party mentions. That leaves the model with incomplete or conflicting signals.

Fix: Align category language across every surface. Update your homepage, product pages, G2 profile, and comparison content to use the same category terms. For Wikipedia, propose changes on the article's talk page, since editors with a conflict of interest are strongly discouraged from editing the article directly.

Timeline: In our experience, category corrections often take hold faster than net-new recall because the model already knows you exist.

How to improve AI brand recall#

Three areas of work address recall and category accuracy: aligning category language, earning third-party validation, and structuring content for passage retrieval.

Align category language#

Category accuracy depends on how consistently your brand appears alongside category-defining terms. If you sell a "sales enablement platform" but the model files you under "CRM," reverse discovery queries are likely to surface CRM vendors instead.

Discovered Labs fixes this by using the same category language across your content, third-party mentions, and structured data. In the brand recall study, Reddit was associated with 13.1 points higher category accuracy. In our experience, repeating the same accurate claim on your site, on Reddit, in industry publications, and in comparison content helps build the association.

Earn third-party validation#

Third-party validation shapes how models see your brand, though the signals are narrower than they look. In Discovered Labs' study, a Wikipedia article was associated with 10.4 points higher recall but 16.7 points lower category accuracy. Customer ratings barely tracked how models described brands.

The CITABLE framework includes third-party validation as a core component. Research on AI search engines backs this up: they show a systematic bias toward earned media, favoring third-party sources over brand-owned pages.

Reddit carries disproportionate weight at retrieval. Discovered Labs' research found Reddit content in roughly 27% of ChatGPT's search slots during query processing. Consistent, accurate mentions in high-trust communities shape what the model finds about your brand when it searches.

Structure content for passage retrieval#

Retrieval is where structure pays off. Research on dense retrievers, which encode passages as vectors rather than counting keyword matches, shows they can outperform BM25 on passage retrieval. Discovered Labs therefore structures each section to answer a single question, leading with the answer in a compact block.

Ranking and citation also overlap less than most teams assume. In University of Toronto research, the domains cited by GPT-based search overlapped with the domains in Google's top 10 by about 20% to just over 50%, depending on the consumer product category tested. Many of the sites AI cites are not the sites that rank, which is why AI search needs its own playbook alongside SEO.

How to measure and benchmark AI brand recall#

Citation rate, mention rate, and share of voice are separate metrics. Citation rate tracks how often your content is cited in AI answers. Mention rate tracks how often your brand is named, cited or not. Share of voice compares your mention rate with competitors for the same query set.

Score each response#

Run priority buyer queries across ChatGPT, Claude, Perplexity, and Gemini. Score each response on three questions: is your brand named, is the category correct, and is a competitor recommended instead.

Discovered Labs' AI visibility tracker covers citation rate, mention rate, and share of voice for retainer clients. We report results in pipeline terms: MQLs, demos, and revenue-adjacent metrics.

AI visibility measurement is probabilistic. In Discovered Labs' study, re-running the same prompts changed the result for 13% of brands. Measure trends across 30-60 days. Track each model separately, because the same question can produce different answers from memory and from live search.

Benchmark against the study#

Discovered Labs' 1,000-brand study provides a baseline. Recall ranged from 51% on Gemma 4 to 78.5% on GPT-5.6 Luna, depending on the model. If a model recalls your brand, you're ahead of the brands it misses. If it doesn't, you have a clear gap to close.

The state of AI visibility 2026 breaks down AI visibility by vertical and ARR band.

Track five metrics#

Use this measurement table to track progress:

Caption: Five metrics for tracking AI brand recall. Citation rate benchmark: 13.9% across 1,200+ B2B brands, per Discovered Labs' State of AI visibility 2026.

How Discovered Labs helps#

Discovered Labs runs SEO and AI search as one organic search operation, with work on both model memory and retrieval. Client results show the approach in pipeline terms.

Gladia grew sales-accepted leads 7x in 4 months, with 93% of AI-referred leads coming from LLM search. incident.io increased organic meetings booked by 22% and improved AI visibility from 38% to 64%. The common thread: CITABLE-structured content and consistent third-party mentions, including Reddit. For incident.io, the work also flagged legacy content that conflicted with its current positioning.

"It's clear that working on AI visibility is as important now as SEO was in the 2010s... I believe early adopters will win the Answer Engine Optimization battle, so it was important for me to find a partner who would help us see results quickly." - Tom W., CMO, incident.io

Brand authority in LLMs is not the same as domain authority in Google. That's why Discovered Labs prioritizes consistent third-party validation alongside link building.

What to do next#

AI brand recall comes down to two questions: does the model know your brand, and does it file you in the right category. Retrieval work on passage structure and third-party consistency moves faster, while parametric recall builds across model releases. Run the direct and reverse query baseline first, then fix whichever layer is broken.

Book a call and we'll tell you honestly whether we're a fit. Our Search Visibility Diagnostic gives you a full diagnosis across Google, ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews, a prioritized 90-day plan, and a 60-minute findings walkthrough with a co-founder. Retainers are month-to-month, with no long-term commitment.

FAQs#

Is my brand in ChatGPT training data?#

You can't access ChatGPT's training data directly, but you can test whether your brand is in the model's parametric memory. Run direct queries ("What is [Your Brand]?") and reverse queries ("What are the best [Category] tools?") with web search turned off. If the model names you without citing sources, your brand is likely in parametric memory. For ongoing tracking, Discovered Labs' AI visibility tracker measures mention rate and share of voice across five AI engines for retainer clients.

How often do LLMs update brand knowledge?#

Parametric memory changes only when a lab releases a new model, so shifts can take a long time to show. Brand mentions earned between two model versions' knowledge cutoffs can enter the newer model's memory. Retrieval-augmented knowledge updates as the web changes, so Discovered Labs treats retrieval as the faster lever.

Can I improve recall without waiting for retraining?#

Not parametric recall itself, but you can raise how often AI answers include your brand. Start with retrieval-augmented knowledge. Structure content for passage retrieval with answer-first sections, block-level organization, and consistent third-party mentions. Dense retrievers outperformed BM25 by 9-19 points in top-20 passage retrieval accuracy, and Discovered Labs reads that as a reason to favor clear, single-question sections over keyword density.

What's a good AI brand recall benchmark for B2B SaaS?#

In Discovered Labs' 1,000-brand study, recall from memory ranged from 51% of brands on Gemma 4 to 78.5% on GPT-5.6 Luna. If a model recalls your brand, you're ahead of the brands it misses. For citation rate, Discovered Labs' State of AI visibility 2026 cites 13.9% across 1,200+ B2B brands, with breakdowns by vertical and ARR band.

Key terms glossary#

AI brand recall: The knowledge an LLM holds about your brand in its parametric memory before retrieving external content. Commonly measured by whether the model names your brand in response to unbranded category prompts or direct brand queries.

Category accuracy: Whether the model assigns your brand to the correct category. Measured by comparing the model's category assignment to your verified category. In Discovered Labs' experience, misclassification can do more damage than absence because it can send buyer intent toward competitors.

Parametric memory: The collective knowledge encoded in a model's internal parameters, shaped by training data and fine-tuning processes. Distinct from retrieval-augmented knowledge, which pulls from external sources at query time.

Reversal curse: The tendency of LLMs trained on a fact in one direction ("A is B") to fail at recalling it in the other direction ("B is A").

BM25: A traditional ranking algorithm used in information retrieval that scores documents based on term frequency. Dense passage retrievers have shown stronger performance on modern retrieval tasks.

Dense retrievers: Neural retrieval systems that encode passages as dense vectors, enabling semantic similarity matching that can outperform keyword-based approaches like BM25.

CITABLE framework: Discovered Labs' proprietary content methodology covering Clear entity and structure, Intent architecture, Third-party validation, Answer grounding, Block-structured for RAG, Latest and consistent, and Entity graph and schema. Full details in the CITABLE framework post.