Treat a free AI visibility score as a claim, not a measurement, until you see the exact question asked, the engine that answered, and its answer text. Those three artifacts let you re-run the check yourself in about five minutes. Everything below is that test, plus what 18 tools publish about their free tier.

The phrases people actually type are plain: on 28 July 2026, Google's own autocomplete completed free ai visibility with checker, tool, audit, tracker, report and score. What almost nobody types is the harder question underneath — is this number real? An AI visibility check is a short test that puts buyer-style questions to AI answer engines and reports whether your brand appears in what they say back. This article is about auditing that report before you act on it, whichever tool produced it.

To ground it, we re-read the public pages of 17 AI-visibility tools on 28 July 2026 and recorded what each one states you get without an account. We did not create accounts, so this is a survey of published claims, not a logged-in walkthrough — and we say so in every cell. Our own tool is the eighteenth row, and it is the only one we can verify from the inside.

What is an AI visibility check, and what does the number actually measure?

An AI visibility check asks a set of buyer questions to named AI engines and counts how often your brand appears in the answers. The number it returns is a presence ratio over that specific question set, on that specific day — not a rank, not a market share, and not a property of your website.

The unit matters more than the scale. A score of 62 does not mean you are the 62nd-best option, and it does not mean 62 percent of buyers see you. In almost every tool on the market it means some variant of "your brand appeared in this fraction of the answers we collected" — so the fraction's denominator is the whole story, and the denominator is exactly what a headline number hides.

That is why the same page can be scored 0 by one tool and 70 by another without either one lying. They asked different questions, of different engines, on different days, and divided by different denominators. Nothing in the two numbers is comparable except the word "visibility".

Why do two AI visibility tools score the same brand differently?

Four things differ silently between tools: the question set, the engine list, the retrieval method and the model version. Change any one and the score moves, even though nothing on your website did. The useful comparison is therefore not which tool scores you highest, but which one tells you enough to reproduce its number.

The question set moves it first. Goodie's free AI Visibility Index organises its view around "Top Prompts" with a "Topic Volume" column — prompts ranked by how much a topic is asked about. Our own free check derives three questions from the offer described on the site being tested. Both are defensible; they are not the same experiment, and a brand can be absent from a high-volume category question while appearing in a narrow use-case one.

The engine list moves it next, and the spread is wide. Rankscale states it tracks "17+ engines including ChatGPT, Perplexity, Claude, Gemini, and Google AI Overviews"; Writesonic's page says "We crawl ChatGPT, Claude, Gemini, Google AI, and 6 more platforms like a real user"; Ahrefs names five surfaces; our own free check deliberately reads two. A brand that is strong in one engine and invisible in another will look completely different depending on how many engines are in the denominator.

The retrieval method moves it in a way almost no buyer thinks to ask about, and at least one vendor makes the argument openly. ZipTie's page states: "Unlike many tools that rely on APIs - which can produce artificial results - ZipTie mimics real user behavior to capture the actual results your customers see." Whether or not you accept that framing, it is a real fork in the road: an API call and a browser session are not guaranteed to return the same answer, and a tool that never tells you which one it used has withheld a load-bearing fact.

Finally, the model behind the engine name moves it, quietly. HubSpot's AEO Grader page is unusually specific here and names its panel as "GPT-5.4 mini (OpenAI), Perplexity, and Gemini (Google)" — a level of disclosure most pages skip. It matters because "ChatGPT" is a product, not a constant. We learned this on our own measurements: across our internal tracking runs the model versions behind two of the engines we read changed underneath us mid-series, which means any before-and-after chart spanning that change is comparing two different instruments.

So the useful question is never "which tool gives the highest score". It is "which tool tells me enough to reproduce its number".

Which four artifacts should a free AI visibility check hand you?

Four artifacts make a score auditable: the exact question, the named engine, the answer text, and the date with the sample size. With all four you can reproduce the check yourself; with three you can partly reproduce it; with fewer you are being asked to take a number on faith.

1. The exact question, word for word. Not a topic, not a keyword, not a "prompt idea" — the literal string that was sent. A topic list tells you what area was probed; only the string lets you send the same thing again.
2. The engine that answered, named — and ideally the model. "AI" is not an engine. "ChatGPT" is an engine but not a model. The more precisely the panel is named, the more of the result you can reconstruct later.
3. The answer text, unedited. A tick mark says your brand appeared. The text says whether you were the recommendation, a footnote, an alternative, or described as something you do not sell. Those are four different business situations behind one identical tick.
4. The date and the sample size. Generative answers move. A result without a date cannot be compared to anything, and a result without a denominator ("6 checks", "50 prompts") cannot be weighed against anything.

Each missing artifact hides a specific failure. Without the question you cannot tell a decisive category query from a trivial one. Without the engine you cannot act, because the levers for a training-data engine and a live-web engine are not the same. Without the answer text you cannot distinguish a recommendation from a passing mention. Without the date you cannot re-measure — and re-measuring is the only thing that converts a score into evidence.

How do you run a five-minute control test yourself?

Ask three buyer questions by hand in two engines, record six yes-or-no results, and divide. That is the whole method every tool automates, and running it once tells you whether a tool's number is in the same universe as reality.

1. Write three questions a buyer would type, not three keywords. One category question ("Which tools are best for [job] in [market]?"), one alternatives question ("What are the best alternatives to [a competitor]?"), one situation question ("Who should I hire to [specific outcome]?"). Write them in the language your buyers use — a market's real questions are asked in that market's language, not translated from English.
2. Open a fresh, logged-out or temporary chat in each engine. A signed-in session carries memory and personalisation, and it will happily quote things back to you that a stranger would never see. If your assistant has a memory feature, switch it off for the test.
3. Paste each question exactly as written, once per engine. Do not rephrase between engines and do not add "and mention [your brand]". The moment you name yourself, the test is over.
4. Record six rows: engine, question, brand named yes or no, and every other brand named. Paste the full answer text into a document. That document is your evidence; the score is only its summary.
5. Divide. Three questions across two engines is six checks, so each one is worth about 16.7 points. Two hits out of six is 33; four out of six is 67. Now you have a number whose denominator you can defend.
6. Diarise a repeat in two to four weeks with the identical question set. Changes to a site need time to be crawled and re-read, so an immediate re-run measures noise, not progress.

Two mistakes void the comparison, and we have made both. The first is editing the question wording between runs: on our own product's early tracking, the question text drifted from one run to the next, which means the earliest and latest scores were measured on partly different questions and cannot honestly be read as a controlled before-and-after. Only the runs that used an identical question set are comparable. The second is running the test signed in — the answer you get is then partly about you, not about your market.

The control test tells you whether a tool's number is plausible. It does not tell you what your market sees. Six checks are a diagnosis, not a survey: no personalisation, no geography, no repetition, and no coverage of the phrasings you did not think of. Treat it as a smoke test for the tool, and as a floor for your own expectations.

What do free AI visibility tools state you get without an account?

Four of the 17 third-party tools we re-read on 28 July 2026 state a free run with no account: Semrush, Ahrefs, SE Ranking and Goodie. One states a free check behind a form (HubSpot), one asks for an email address (Am I on AI), nine state an account or a trial for the tracking itself, and two are demo-only. Only Semrush's page states that both the prompts and the AI's text answers are shown in full; Goodie's publishes a prompts table but says nothing about answer text.

Read the table for what it is. Every cell records what the vendor's own public page stated on 28 July 2026 — nothing more. We did not create accounts, so "not stated on page" means the page does not make the claim; it never means the product lacks the feature. Where a page uses a phrase that decides the cell, we quote it. The Webappski row is different in kind, and we flag it rather than hide it: we can read our own production code and our own live page, so that row is verified from the inside while every other row is verified from the outside.

The selection rule, so you can rebuild the list yourself: every tool named in Zapier's roundup "The 8 best AI visibility tools in 2026", plus every AI-visibility checker an English-speaking buyer lands on from the phrases Google autocompletes for "free ai visibility". Similarweb appears in that roundup but we found no public product page at the URLs we tried on 28 July 2026, so it is absent rather than assessed. Self-hosted open-source trackers are excluded because they offer no hosted free check for a visitor to click. Tools sold specifically into a single non-English market are excluded too: an English-language pool should be judged against the tools an international buyer actually meets.

One caveat we owe you, because it cuts against us. In a hands-on run of our own on 24 July 2026, Semrush's free result did show prompts word for word and a text preview of the answer, while the button that opened the full answer led to a registration prompt. We did not repeat that hands-on run on 28 July, so the table above records only what the page states today. Read the Semrush row as the strongest published claim in the set, not as a defeated competitor.

The pattern across the eighteen rows is less about generosity and more about vocabulary. "Free" is used for at least four different things: free and instant with no identifier, free after a form, free after an account, and free as a trial of a paid product. Before you compare scores, check which of the four you were given.

What does the Webappski free check show, and where does it stop?

Our free check asks three buyer questions of ChatGPT and Gemini — six checks in total. With no account and no email it shows the presence score, which engines named you, all three questions in full, one complete answer in the engine's own words, and the rival brands those answers named. The other five answers, the sources and the recommendations need an email address.

The mechanics are deliberately dull. You enter a domain and a brand name at the free AI visibility check; the questions are generated from what the site says it sells, in the language the site is written in; the result is a three-by-two matrix in which every cell says whether the brand was named. One cell is opened all the way — the full text of that engine's reply — so at least one number on the page is checkable rather than merely asserted.

The rival-brand list has its own honesty contract, because a language model can invent a company as easily as it can name one. Every answer is read by two independent extractor models. A name is listed only if it appears literally in the engine's own text, and your own brand is filtered out. If both models name a brand it goes in the main list; if only one does, it goes in a separate, explicitly weaker list, and the two are never merged — merging would quietly turn a guess into a confirmation. If no rival was named at all, the result says so, because "we looked and found nobody" and "we did not look" are different facts and must not render the same way.

Now the limits, in the same detail. Five of the six answers and all the citations sit behind an email address — that is a real gate, not a rounding error. Rival names came into the free view on 28 July 2026, so a domain whose result is being replayed from a check run before that date shows no rival block at all rather than an empty one. Six checks are a starting point, not a market model: no personalisation, no logged-in session, no third engine, and no guarantee that the same six checks tomorrow return the same six answers. A score of 0 means the brand was not named in these six checks — nothing wider than that. If you want the reasoning behind a fuller audit, we set it out in how we check whether AI recommends a site, and we wrote about the trustworthiness of visibility scores in general in can you trust an AI visibility score.

Common questions about free AI visibility checkers

Most questions about free AI visibility checkers reduce to four. Is the number accurate, how many questions make it meaningful, why does it change between people and days, and what does the email address actually buy. Short answers follow, with the limits attached.

Are free AI visibility checkers accurate?

They are accurate about what they measured and often silent about what that was. A free check is a small, dated sample of AI answers, so the honest reading is "this is what these engines said to these questions on this day". Accuracy problems almost never come from arithmetic; they come from an undisclosed question set, an unnamed engine, or a simulated result presented in the visual language of a measured one.

How many prompts do you need before an AI visibility score means anything?

More than a free check gives you, and the industry knows it — Otterly publishes a calculator whose stated purpose is "to determine how many prompts you need to monitor to get statistically reliable" results. Six checks, as in our own free tool, are a diagnostic: enough to tell "absent everywhere" from "present somewhere", not enough to support a budget decision. Treat small samples as a direction and repeated, fixed question sets as evidence.

Why does my brand appear in ChatGPT for one person and not another?

Because a generative answer is not a stored ranking. A signed-in session carries memory and personalisation, phrasing differs between users, and the same question can return different text on two consecutive runs. This is precisely why a control test uses a fresh, logged-out session with a fixed wording, and why any single result — ours included — is a snapshot rather than a position.

Can I check my AI visibility without giving an email address?

Yes, in several tools, though "without an email" and "without an account" are not the same promise. On 28 July 2026 the public pages of Semrush, Ahrefs, SE Ranking and Goodie stated a free run with no sign-up, HubSpot's stated a free check behind a form, and Am I on AI's form asked for an email. Our own check runs on a domain and brand name alone; the email gate begins at the fifth of six answers.

Is an AI visibility score the same as a search ranking?

No. A ranking is a position in an ordered list that exists whether or not you look at it. An AI visibility score is a presence ratio computed from answers that were generated at the moment of asking and may never be generated the same way again. Any tool that presents its output as a stable rank is describing a different thing from the one it measured.

How often should I re-run the check?

Every two to four weeks, with the question set unchanged. Anything faster measures the natural variation between runs plus the delay before an engine has re-read your site; anything slower and you cannot tell which change moved the number. Keeping the wording frozen matters more than the frequency — a re-run with new questions is a new experiment, not a comparison.

So what should you do with a free AI visibility score?

Use it as a hypothesis, then spend five minutes disproving it. If the tool showed you the question, the engine and the answer text, you can reproduce the check by hand; if it did not, run the control test first and let the score argue with your own six results.

A short decision guide. If you want a fast, checkable read with the questions and one full answer in front of you, run a free instant check and read the answer, not just the number. If you want competitor brands with their own visibility figures attached, Goodie's free index publishes those and ours does not. If you are buying continuous monitoring, judge Peec, Profound, Writesonic, Rankscale and the rest on a demo against the four artifacts above — not on fields that stay invisible in the public version. And if you want to see where you stand before deciding anything, do the control test: it costs nothing but attention and it is the only version of the result you fully own.

If you would rather have it done for you in about half a minute, you can run our free AI visibility check without an account. You will see the three questions, which of the two engines named your brand, one complete answer in the engine's own words, and who was named instead of you — before you decide whether the rest is worth an email address. Whatever tool you use, keep the same standard: a number you cannot reproduce is an opinion with a decimal point.

Method note. Published 28 July 2026. The comparison records what each vendor's own public page stated on 28 July 2026 for 17 third-party tools; we did not create accounts, so cells read "not stated on page" rather than "absent from the product", and demo-only products are marked as not verifiable without a demo. The Webappski row is verified against our own production code and live page on the same date and is therefore evidenced differently from the rest — we disclose that asymmetry rather than smooth it over. Webappski publishes this comparison as the vendor of one of the tools in it. Where an earlier hands-on observation of ours conflicts with a page's current wording, both are stated with their dates. Product boundaries change; re-check before quoting.