/ The Bottom Line Up Front

A model can know your name and still describe you as the wrong kind of company.

Recognition and correct description are two different measurements. In the 1,000-brand study behind this tool, 510 brands were volunteered unprompted and 194 were filed under the wrong category. Zoho is both: named readily, read well, and ranked as CI/CD rather than CRM.

What the ModelRank brand audit measures

The ModelRank brand audit runs four probes against an open model and reports what each one found. It does not ask the model a question and grade the answer. It reads the model directly, which is why the result is stable enough to re-run and compare rather than being a fresh sample of a stochastic chat.

The first probe asks what exists in your category and counts how often your brand is volunteered without being named first. The second ranks your true category against four length matched distractors on mean per-token logprob, so a shorter label cannot win for being shorter. The third reads twelve buying dimensions out of activation space. The fourth follows your brand name up the layers to see how far it surfaces.

Phase two then fans out across you and every competitor you name, measuring the third-party surfaces, the backlink footprint and the own-property facts that the study associated with recall and with factual accuracy. Every comparison there is an ordinal placing among the brands you named, never a percentile against a cohort you are not part of.

Where a probe cannot run, the audit says which probe, and why, and shows nothing else. An unavailable measurement is reported as unavailable rather than as a zero, because a zero on a chart is a claim and a missing measurement is not.

Four probes, each reported on its own terms

Each probe needs a specific capability from the serving endpoint. When one is missing, the report names the probe and the reason and draws nothing in its place.

Recall

Named unprompted

The model is asked what exists in your category, several times over several prompt frames. We count the generations that name you on a word boundary and report the rate with its sampling interval.

Category binding

What the model thinks you sell

Your true category is ranked against four length matched distractors on mean per-token logprob. The margin is the gap to the runner-up, in nats per token.

Sentiment

How the model reads you

Twelve buying dimensions read from activation space rather than from generated text: ease of use, pricing value, support, reliability, documentation, onboarding, integrations, product depth, innovation, trust and security, community, trajectory.

Logit lens

Where the name surfaces

The brand name is followed up the stack layer by layer. A name that peaks mid stack and fades is held differently from one that resolves at the top.

How the ModelRank brand audit works

01

Name your domain

We read a handful of your pages to place you in the category taxonomy, and you can name up to seven competitors or let the crawl suggest them.

02

Four probes run live

Recall, category binding, activation-space sentiment and the logit lens. Each one streams back the moment it lands, so the report fills in as it is measured.

03

The channel phase queues

Third-party surfaces, backlinks and own-property facts for you and every named competitor, measured per brand and placed as an ordinal among the peers you named.

04

The audit keeps a link

Every run is stored under its own key, so the finished report can be reopened and shared without running the model again.

A worked example: Zoho

The model names Zoho readily and reads the product well on eleven of twelve buying dimensions. It still ranks CI/CD above CRM when asked what Zoho is, by 0.36 nats per token. Recognition and correct description are separate measurements, and a brand can pass one while failing the other.

Example report for Zoho

Zoho ranks 2 of 12 measured CRM brands on gemma-4-12b, against 4 named competitors. The model volunteers the name, reads the product well on every buying dimension, and still files it under the wrong category: a brand can be well known to a model and still be described to buyers as the wrong kind of company.

Read the Zoho report

This is Zoho, not you. Every number on that page is read out of the published 1,000-brand ModelRank study on gemma-4-12b. Your own audit measures your domain live, against the same model, and opens on a page of exactly the same shape.

Explore our other tools

More free tools to help optimize your content for AI search.

Reddit threads finder

Find threads cited in ChatGPT and Gemini for AI visibility and AEO

Try it

AEO content evaluator

Grade your content for AI search optimization and citation readiness

Try it

Headline optimizer

Generate and rank headline variants for AI search and SEO

Try it

LLM eval calculator

Calculate optimal sample sizes for AI evaluation experiments

Try it

AI Assist Widget

Add AI buttons to any website to boost AEO and brand citations

Try it

Agentic browsing checker

Run the Lighthouse agentic-browsing audit: WebMCP tools, schemas, llms.txt, CLS

Try it

Frequently asked questions

What is a ModelRank brand audit?

It is a direct measurement of what an open weights model holds about a brand: whether the brand is volunteered when the model is asked what exists in its category, which category the model ranks highest for it, how the model reads it on twelve buying dimensions, and how far up the layers the brand name surfaces. It reads the model rather than grading a chat answer.

How is this different from asking ChatGPT about my brand?

Asking a chat assistant gives you one sample of a stochastic generation, shaped by the system prompt, the retrieval layer and the conversation so far. The audit measures the weights instead: token probabilities, activations and layer-by-layer resolution. That makes the result reproducible and comparable between runs, which a chat answer is not.

Which model do you measure against?

An open weights model served on our own endpoint, so the measurement can be reproduced and the layers can be read. The audit names the exact model it ran against at the top of every report. Nothing here is a measurement of a closed commercial assistant, and the report never claims to be.

What does it mean if my brand reads as deviated?

It means the model ranks another category above your real one when asked what your brand is. In the 1,000-brand study behind this tool, 194 brands in 1,000 were deviated. Answers about a deviated brand start from the wrong category, so a buyer question gets framed against the wrong set of alternatives before the model gets to your product.

Why do some probes come back unavailable?

Each probe needs a specific capability from the serving endpoint: a generation route for recall, scored continuations for category binding, the stored sentiment direction vectors for the sentiment read, and a logit lens route for the depth read. When one of those is missing the audit reports which probe could not run and why, and draws nothing in its place. It never fills the gap with an estimate.

What is the channel phase?

Phase two of the audit. It fans out across you and every competitor you name, measuring third-party surfaces, the backlink footprint and the facts a crawl can establish about your own pages. Each metric is reported as an ordinal placing among the brands you named, because at four to eight peers a percentile would be a false precision.

Can I compare myself against named competitors?

Yes, in the channel phase. You can name up to seven competitor domains, and the audit measures each one on the same channels you are measured on. The model phase is measured for your brand only, since a model measurement of a peer would need its own consent and its own run.

Is the report shareable?

Every audit is stored under its own key and has a permanent link. Reopening the link renders the stored audit, including whichever part of the channel phase has finished, without running the model again.

How long does the audit take?

The model phase usually finishes in one to three minutes, longer when the serving endpoint has scaled to zero and has to start cold. The channel phase is queued behind it and keeps running after you close the tab, which is what the stored link is for.

Is the tool free?

Yes. Enter your domain and your work email and the audit runs. The email is what turns an anonymous request into a report we can send you a link to, and it is the same gate the other free tools on this site use.