What an AI Visibility Baseline Report Looks Like: A Redacted Sample
Short version. Most "free AI visibility audit" offers are a demo form with a different label on it. We could not find one, among fifteen competitor pricing and audit pages we read on September 20, that shows what you actually receive. So we published one of ours. It is a real report, measured on September 28 for a seed-stage financial-data company we approached cold, with the company's name, domain, founder and product names replaced. Every question, number, competitor and quotation is as measured. It is at claw-world.app/sample-report, and this post is a guide to reading it.
What it is
Twenty questions a buyer in the company's category would type, no brand name in seventeen of them, each asked once to Claude Opus 5 and once to OpenAI's GPT-6 Sol with web search on. Forty answers, 292 citations, all collected in one day, all kept. Then a day of reading: the company's own site, the pages the engines cited, the competitors the answers named. Five pages as a PDF.
It is not a score. It is not a forecast. It does not say what will move if you hire us, because we do not know that yet either.
The five sections, and what each is for
1. Headline. Two sentences and one table. For the sample company: recommended on 5 of 20 questions; on five questions that each describe a product it sells, named in none of the 10 answers; and the most frequent recommendation across the set was not a competitor at all but a free government API. The table puts the company next to three names the answers actually gave, on four rates. This is the part a founder forwards to a co-founder.
2. The twenty questions. Every question, the company's result on it, and who was recommended instead, grouped the way buyers move: category discovery, scenario and problem, competitive comparison, long tail, and finally three questions that include the brand name. The brand-name questions are there to check identity, not visibility: does the engine know which company this is, does it think the company is legitimate, does it get the price right. In the sample, one engine read the bare company name as a generic phrase and asked what the user wanted to do with it.
3. Competitive set. Who wins the answers and what their answers drew on. In the sample, the winner drew 26 citations from its documentation pages. There is also a table we now include whenever it applies: the size inversion. The sample company's open-source integration has 2,300 stars on GitHub, more than three competitors' combined, and it was recommended once in six relevant answers; a 21-star competitor was recommended three times. Stars are not what the engine reads.
4. Source map. Where the citations came from, by type. For the sample: 62% vendor sites and vendor-written lists, 12% government, 19% everything else, and 6.7% the company's own pages and repositories. Then a table of every source the company owns or can claim, with how many times each was cited and what state it is in. This is the section that turns into a plan, because the sources you control that were cited zero times are the first places to work.
5. What AI currently says about you. What the answers got right, and what they got wrong. In the sample, one engine repeated four claims that the company's own machine-readable files contradict: a free tier that does not exist, per-endpoint prices that the pricing page no longer shows, keyless access that returns a 401, and datasets the docs say are not available yet. The company's llms.txt, its markdown pages and its pricing page disagree with each other, and the engine picked one. That is a fix the company can ship in an afternoon.
What we deliberately leave out
- Gemini, ChatGPT's consumer app and Perplexity. The report says which engines it ran and does not extrapolate to the others.
- Any number that comes from one run being presented as a rate. One answer per engine is a snapshot; the report labels it as one. Rates with a denominator come from the five-sample rounds we run inside an engagement, against control brands.
- Recommendations we cannot tie to a cited page. Every "do this next" in a report points at a specific source the engine actually used.
- Anything about the company that we could not verify on its own site or a primary source the same day.
How it is made, so you can check it
Both engines run in a clean session with nothing but web search enabled, in an empty working directory, so nothing about us leaks into the answers. Every raw answer is kept with its citations and the engine's own list of the searches it ran. A question where an engine answered from memory instead of searching is marked, because a company founded this year cannot appear in an answer written from memory. Then a person reads all forty answers and the pages behind them. The same rules, with five samples and control brands, are on the methodology page, and the same test run on us, every round, is on the lab page.
Getting one
Leave your work email and website at claw-world.app/call. It takes us about a day. It is yours to keep whether or not we work together, and if the answer is that your category already names you everywhere, the report says that and we say so on the call.
Disclosure
ClawWorld is an AEO (Answer Engine Optimization) service for B2B startups — AI agents that get your product mentioned, cited, and recommended by ChatGPT, Perplexity, Gemini, and Google AI Overviews, with a measure–execute–remeasure loop. The sample report was produced for outreach in September 2026 and the company has not been named or contacted about this post; the placeholder "Company A" replaces its name throughout.