TL;DR
- A buyer shopping for SEO or GEO help can now meet two things: a free tool that scores their "AI visibility", and an AI answer that names the best agencies, drawn in part from "top 10" lists. Before either counts as evidence, ask what was measured and who wrote it.
- "AI visibility score" can cover three different instruments: a page audit, a model's opinion of your site, and a count of how often AI answers name you. Only the third measures visibility. The first two measure your website.
- A measurement you can check states six things: the questions, the engines, the date, the number of runs, whether it counted mentions or citations, and where the raw answers are. Without them, a number cannot be re-run or compared.
- A "top 10 agencies" list published by the agency ranked first is an advertisement laid out as a ranking. At least one agency has said in public that the tactic gets it cited by AI assistants, and that it expects the effect to be temporary.
- We sell audits and compete with agencies that offer free scores and publish these lists, so we have an interest in this subject. The checks below apply to our numbers too.
Ask ChatGPT or Google's AI Mode for the best SEO agency in your city and you get names. An agency may also offer to score your own AI visibility for free, in a few minutes. Both look like evidence. This post is about how to read them as a buyer. It names no other agency and no scoring tool, because the checks are the same for all of them, ours included.
A word on the sources first. The research figures come from an October 2026 preprint from the University of Toronto (not peer reviewed, English questions about US markets), an Ahrefs benchmark we covered in an earlier post (vendor data, English and mostly US), and our own small probe on Quebec roofing in July 2026. The statement about self-published lists is a single agency's public account of its own tactic. Nobody has measured it. Google and the Competition Bureau are quoted from their own pages.
Why do free AI scores and top 10 lists show up together?
Free AI visibility scores and top 10 agency lists both reach a buyer at the moment of choosing an agency, and both are cheap to produce. A list page answers the exact question a buyer types. An instant score gives an agency a number and a reason to start a conversation.
Lists were also a large share of what ChatGPT drew on for early-stage buying questions in one vendor benchmark. In Ahrefs' AI Search Benchmark for Q1 2026, "Best X" blog lists were 43.8 percent of ChatGPT's source URLs across 750 top-of-funnel queries, and 40.6 percent in the agency vertical (we went through that study in the AI visibility funnel). Ahrefs sells AI visibility tracking, and its data is English and mostly US. On that evidence, for top-of-funnel questions in the agency vertical, about four in ten of the source URLs ChatGPT drew on were lists, and somebody wrote each one.
A free score is a fair way to open a sale. We sell audits ourselves, and a first look at a site is a reasonable place to start. What needs checking is what the number is said to mean.
What does a free AI visibility score actually measure?
A free AI visibility score can come from three different instruments, and two of the three measure your website, not how often AI answers name you. When a tool hands you a figure out of 100 in a few minutes, with a handful of sub-scores, the first job is to work out which instrument produced it.
| Instrument | What it does | What it can tell you | What it cannot tell you |
|---|---|---|---|
| Page audit | A crawler reads your pages and checks features: crawler access, indexing, structured data, headings, speed | Whether search engines and AI crawlers can reach and read your pages. A blocked crawler or an unindexed page is a real finding | Whether any AI answer names you |
| Model opinion | A script hands your URL or your name to one or more AI models and asks them to rate or describe it | What a model says about your site once it is pointed at it, including errors in what it says | Whether an engine would find you when nobody hands it your URL |
| Answer measurement | A script sends a fixed list of buyer questions to named engines, several times, and counts who is named and which pages are cited | How often you and named competitors appear in the answers a buyer would see, per engine, on that date | Why you are or are not named. The count also goes out of date |
The first two are legitimate work. A page audit is most of what an SEO audit is, and ours contain one. What does not hold up is calling either one a visibility score. A study by Mahe Chen and Nick Koudas of the University of Toronto (arXiv 2610.15033, version 1, 10 October 2026) collected 90 days of answers from ChatGPT, Gemini and Perplexity to commercial questions, and then replayed the questions through the ChatGPT and Gemini APIs to see what those two engines searched for. Across pages of the same website, page features explained 2.8 percent of the variation in how often ChatGPT cited them. The engine's own retrieval explained 32 percent. Every one of the 22,315 pages ChatGPT cited in the replay was a page its own search had returned.
In the same data, FAQ blocks, schema.org markup, summary boxes and author boxes showed no significant effect on citations once pages of the same website were compared. Those are associations, measured on English questions about US markets, through small API models and not the consumer apps. Google's own guidance says structured data is not required for its AI features, which we covered in how AI chooses what to cite. In the Toronto data, then, a score built from page features grades the part that explained about 3 percent of ChatGPT's citation differences within a website.
A model's opinion of your site has a different gap. Handing an engine your URL skips retrieval, the step where the engine decides which pages to read at all. A favourable description of a page the engine was given says little about whether it would have found that page for a buyer's question.
A score out of 100 is not the problem in itself. Our Local SEO Audit includes a 0 to 100 health score across seven weighted categories. It grades the SEO health of a site and sits beside a ranking table against named competitors. A composite number is fine when the report says what went into it and what it stands for.
We have put that kind of number on a client's site ourselves. One of our case studies reports an April 2026 audit that scored a site 55 out of 100 for AI readiness. That score graded the site, so by the table above it is not a visibility measurement. The visibility reading in the same case study came from a different instrument: four buyer questions put to Perplexity and Gemini with web search on, in July and again in September 2026.
How do you check an AI visibility score?
An AI visibility measurement you can check states six things: the questions, the engines, the date, the number of runs, what was counted, and where the raw answers are. Each of the six changes the number, so a figure that leaves one out cannot be re-run by you or compared with anyone else's.
| Ask for | Why it changes the number |
|---|---|
| The exact questions | A page is cited for particular questions, not in general. The list should be written the way a buyer asks, in the language your buyers use, without your name in the question. With no list, the score cannot be re-run |
| The engines, by name and surface | ChatGPT, Gemini, Perplexity and Google's AI Overviews and AI Mode are separate systems. Ahrefs found AI Mode and AI Overviews shared 13.7% of cited URLs for the same queries, while reaching the same conclusion 86% of the time. App or API, search on or off, also matter: in our roofing probe, with search off, 7.1% of responses named a local business our automated check matched to a real one (5 of 70), and 18.9% once we counted one engine that retrieved live results anyway (17 of 90) |
| The date | The cited pages turn over. In the Toronto data, ChatGPT's citation counts ten weeks apart shared 14% of their within-website variation |
| The number of runs | In the Toronto replay, ChatGPT and Gemini rewrote the question into their own searches on every run. Two same-day ChatGPT runs overlapped by 2.7% on search queries and 15% on cited pages. A single run does not tell you what the next one will return |
| Mentions or citations | Being named in the answer and being linked under it are separate counts. In the Toronto data, 46% of ChatGPT's mentions of a brand the question did not name had no cited page naming that brand |
| The raw answers | The log lets you verify the count and catch invented names. Our own probe returned at least one company name that was garbled or made up |
None of this needs a platform. The ten-minute version is in our post on what AI answers from memory, and the five-stage version is in the AI visibility funnel. A number that comes with all six can serve as a baseline. One that comes without them cannot be checked. Ask us for the same six.
Who wrote the top 10 list the AI is quoting?
The top 10 list an AI answer quotes can be written by the agency ranked first on it, so open the citation and check the publisher. A "top 10 agencies" page published on an agency's own site, with that agency in first place, is an advertisement laid out as a ranking.
We know the tactic is in use because at least one agency has said so in public. In a public interview, it described publishing a top 10 list of agencies on its own site, with itself in first place, and being cited by AI assistants as a result. It added that it expects the effect to be temporary. We are not naming the firm, and a single account is not a measurement.
The account fits what has been measured about how an engine names a business it does not already know. In the niche market of the Toronto preprint (proposal software for architecture, engineering and construction firms, 55 questions), a cited third-party page that named a vendor went with 39.7 times the odds that ChatGPT's answer named it. We went through that result in the rich get richer, the fit get rich. The study counted third-party pages. It did not test a page a vendor wrote about itself, so whether a self-published list earns the same lift has not been measured. Our inference is narrower: an engine reading a list has no dependable way to know that the author and the first entry are the same company, unless the page says so.
Reading a list takes about two minutes:
- Open the citation. Whose website is it on? If the publisher is also an entry, you are reading an ad.
- Look for the method. A ranking should say what was assessed, by whom, and when. A list with no stated criteria is the opinion of whoever published it.
- Look for disclosure. A list can sell its slots. A paid slot is a placement, and the page should say so.
- Read the other entries. Facts you can verify about each firm suggest someone did the work. One flattering line each suggests filler around the first entry.
- Ask again. Put the same question to a second engine, and to the first one a week later. Count a name only if it comes back.
Even an honest ranking is thin evidence for a buyer. Dated results on a real client, with the source and the window for each number, tell you more. Ours are on our case studies page, including what did not move.
Does ranking yourself first on your own list work?
Ranking yourself first on your own list can get you cited by AI assistants for now, by one agency's public account, and nothing we have read measures how long that lasts. There are three reasons not to reward it as a buyer, or copy it as a business.
- Google's guidance points away from it. Its guidance on AI features says: "seeking inauthentic 'mentions' across the web isn't as helpful as it might seem." Google does not mention self-ranked lists there. Treating one as an inauthentic mention is our reading. Google adds that its generative AI features depend on its core ranking systems and on the systems that block spam. That speaks for Google's surfaces only. We have not found comparable guidance from OpenAI or Perplexity.
- In one lab benchmark, manipulation aimed at AI rankers was detectable in principle. In the preprint we covered in GEO tricks, measured, an AI judge separated every method that produced a notable gain from unedited text, to varying degrees, although a filter run at one fixed threshold caught far fewer. The test used one judge model and small open rankers plus GPT-4o-mini, not ChatGPT search, Perplexity or Google's AI features as deployed. It tested rewrites of a page, not self-ranked lists, so it says nothing direct about lists.
- In Canada, a ranking is also a representation. The Competition Bureau's page on false or misleading representations reads: "All representations, in any form whatsoever, that are false or misleading in a material respect are subject to the Act." It adds: "To determine whether a representation is false or misleading, the courts consider the 'general impression' it conveys, as well as its literal meaning." A page that reads as an independent ranking and is written by the firm in first place is worth weighing against that sentence. We are not lawyers, and we are not saying of any particular list that it breaks the law.
For a buyer, a self-ranked list is a sample of how its author handles evidence when its own sales are involved. Ask whether the same tactic would be used in your name, because your name would then be the one attached to it.
What does this not tell you?
This post does not tell you that any particular free tool is wrong, or how common self-ranked lists are.
- That a given free score is worthless. We have not tested the internals of any tool. Some may run real answer measurements behind the number. The six questions tell you which kind you have.
- How often AI cites self-published lists. We have a single public account and no count.
- Anything measured on French-language Quebec questions, apart from our own probe, which was one trade, twelve French prompts and nine models, run with search off (one engine retrieved anyway). The Toronto figures are US and English, associations and not effects, from engines that change continuously.
- That a checkable measurement is a good result. It only tells you the number can be re-run. A number that can be re-run can still be zero.
- What the law says about a specific page. That is a question for a lawyer.
What we do with it
Three things follow for our own work. None of them guarantees a result.
- We do not publish a ranking of agencies. You will not find us in first place on a list we wrote.
- We measure answers on named engines, and re-test. In our heavy civil contractor case study, we put the same four buyer questions to Perplexity and Gemini with web search on in July, August and September 2026, and recorded which answers named the company and which pages they cited. It is a small reading: four questions, and four answers in the July test. That July test was the first, and it ran after the pages had been rebuilt in April. Our own lesson from the mandate, written into the case study, was to run it before the work starts, so the baseline exists. Our Local SEO Audit includes a review of how a business shows up in AI answers such as AI Overviews.
- Run it yourself before you buy from anyone. Five buyer questions, three engines, two runs a week apart, written down. Then hold any score, and any list, against your own log.
Before you hire on the strength of a score or a list, ask who measured what, on which date, and whether you could run it again.