How to check your AI visibility for free: the method, and what the result is actually worth

You can check your AI visibility for free, today, with nothing but a browser. What you cannot do is check it in five minutes, because a single answer from a single assistant is noise. In a study of 2,961 runs across ChatGPT, Claude and Google’s AI, the chance that the same prompt returned the same list of brands in any two responses was below one in a hundred for ChatGPT and Google’s AI, and only slightly better for Claude. So the method matters more than the tool: ask buying questions rather than brand questions, ask each one several times, ask in more than one assistant, use a clean session, and write down who was named and what was cited.

Done that way, an afternoon gives you a defensible picture. Done the usual way, it gives you a comforting one.

The manual method below costs an afternoon and tells you where you stand. It does not move the number, and neither does any dashboard. What moves it is a complete, sourced answer to each of those buying questions, published on your own site and kept current. That is what scaile runs as managed infrastructure: a named AI Search Strategist, research from your own knowledge base, every fact checked, your team approving before anything publishes, live in 14 days. If you would rather not spend the afternoon, the free AI Visibility Check runs the same idea automatically.

Why does asking ChatGPT about your own brand prove nothing?

The most common free check is to open ChatGPT and type “what do you know about Meyer Verpackungen” or “is Meyer Verpackungen a good supplier”. Both are leading questions. You have handed the assistant the entity, so it will retrieve pages about that entity and describe it, usually kindly. You learn that a company by that name exists on the web. You do not learn whether anyone would ever have arrived at you.

The questions that matter are the ones a buyer types before they have a shortlist. Compare the two forms:

• Brand question: “Is scaile a good GEO agency?” Useless. The answer is about scaile because you made it about scaile.
• Buying question: “Which agency should a German mid-sized manufacturer hire to get mentioned in ChatGPT?” Useful. Now the assistant has to choose, and the names it chooses are your real competitive set.

Write ten of these, in the language and market your buyers actually use, because a German buyer asking in German gets a different retrieval than the English translation of the same question. Cover the four shapes that carry commercial weight: the category question (“best X for Y”), the alternatives question (“alternatives to ”), the criteria question (“what should I look for when choosing X”), and the problem question, phrased the way somebody describes a problem before they know the product category exists. If ten is hard to produce, that difficulty is itself the finding, and it is the same gap the platform closes when it derives the questions from your own offering and buyer.

How many times do you have to ask before the number means anything?

More often than feels reasonable. SparkToro and Gumshoe had 600 volunteers run 12 prompts through ChatGPT, Claude and Google’s AI a combined 2,961 times in November and December 2025. For ChatGPT and Google’s AI there was less than a 1 in 100 chance that two runs of the same prompt returned the same list of brands, and Claude was only slightly more likely to repeat a list. Order was worse still: roughly a 1 in 1,000 chance of the same list in the same order. The study’s own author suggests asking a prompt at least 60 to 100 times and averaging the answers, which is also how often each prompt was run in the study itself.

You are not going to run 100 times by hand, and you do not need to. You do need to stop treating one run as a result. Five runs per question per assistant is the realistic manual floor, enough to separate “named in five out of five” from “named once”. Anything you saw exactly once is a coin flip, not a finding.

Record it as a fraction, not a yes or no: named in 4 of 5 runs, competitor A in 5 of 5, competitor B in 2 of 5. That fraction is the only manual number that survives contact with the instability.

Why is one assistant not enough?

Because they are not reading the same web. Ahrefs compared the 50 most-mentioned sources in ChatGPT, Perplexity and Google AI across roughly 76.7 million AI Overviews and about 950,000 prompts each in ChatGPT and Perplexity. Only 7 of those 50 domains appeared in all three lists. The other 43 appeared in one or two of the lists but not all three.

The spread shows up against Google’s ranking too. Across 15,000 long-tail queries, only 12 percent of URLs cited by AI assistants ranked in Google’s top 10 for the original prompt: Perplexity pulled 28.6 percent of its citations from the top 10, ChatGPT between 6 and 8 percent, Gemini 8.6 percent. Google’s own AI Overviews sit apart from all three, on a separate dataset again: Ahrefs’ research into AI Overview citations put the share of cited pages ranking in Google’s top 10 at around 76 percent, revised down to 38 percent in its March 2026 update. Even at 38 percent that is far above the assistants, which suggests AI Overviews function as an extension of the index rather than as a separate retrieval system.

Practically: run every question in ChatGPT, in Gemini or Google’s AI Mode, and in Perplexity. If you only have time for two, take ChatGPT and Google, because they behave the least alike.

Why does the session have to be clean?

Because the assistant knows you. ChatGPT uses saved memories and past chats to personalise responses, which means the account that has spent six months discussing your own company is the worst possible instrument for measuring whether strangers hear about it. OpenAI’s own documentation is the fix: a Temporary Chat neither uses existing memories nor creates new ones. Use that, or a logged-out session, or a fresh browser profile.

Two more controls are worth the seconds they cost. Set the location, because assistants localise and a check run from your office answers a different question than one run from your buyer’s city. And give each question its own conversation, or answer two will be contaminated by answer one.

What exactly should you write down?

Not “did we appear”. Appearing is not the variable that moves. Build a table with one row per question and record four things.

Who was named, in order. Position one in a three-name answer is a different commercial outcome than position five in a list of eight, and the order is the least stable part of the answer, which makes it the most informative when it is stable.

How many names the answer contained. An answer that names twelve providers is a directory. An answer that names three is a shortlist, and only the shortlist matters.

Which sources were cited. Copy the actual URLs.

Whether the assistant searched at all. ChatGPT shows when it has browsed. An answer produced without a search is coming from training data, and no change to your website will move it before the next model generation.

After ten questions across three assistants you have a leaderboard, and the leaderboard is the deliverable. Not “we are invisible”, but “we are named in 3 of 30 runs, the competitor we never lose deals to is named in 22, and in 18 of those the cited page is a comparison article neither of us wrote”.

Why do the citations matter more than the mention?

Because they tell you where the answer came from, and it is usually not the brand’s own site. In an analysis of 240 brand-related prompts producing 23,387 citations across five platforms, 23 percent of citations came from the brand’s own domain, 48 percent from earned sources such as editorial, forums, review sites and directories, and 30 percent from other companies’ commercial content.

That last number is the one people miss. Roughly a third of what the assistant reads about a named brand is published by other companies, often competitors, though the bucket covers any brand-authored commercial content and not only direct rivals. Note the scope: these were prompts containing a brand name, not the open buying questions the rest of this page tells you to ask. When an assistant recommends a competitor and cites that competitor’s own comparison page, you are not losing on product, you are losing because they wrote the page that answers the question and you did not.

So read the cited URLs before you conclude anything. They are the shortest route to a plan, and they are the part a screenshot of the answer throws away. For the mechanics behind which passage gets lifted, the overview of the levers with evidence sorts them by what the research actually supports.

What do you do with a bad result?

Three different problems produce the same symptom, and they have three different fixes. Separate them before you spend anything.

You cannot be read. The assistants never fetch your pages. Check whether OAI-SearchBot, GPTBot and ChatGPT-User are allowed, then check your server logs for real hits from them, because a common failure is a robots.txt that permits them and a CDN that blocks them by IP range. For Google’s AI features the stated requirement is that the page is indexed and eligible for a snippet, and Google says explicitly that no new machine-readable files or special markup are needed. The free AI Search Health Check runs the robots.txt half of this for you, 103 checks across five areas: technical foundation, content architecture, AI optimisation, meta and social, and experience and trust. It reads robots.txt for GPTBot, ClaudeBot, PerplexityBot and CCBot; it does not cover ChatGPT-User and it cannot see your server logs, so the log check stays a manual step.

Nothing on your site is worth quoting. You are reachable, and there is still no passage that answers the question completely, with evidence, in full sentences. A benefits-led product page cannot be lifted into an answer. This is the finding no setting fixes, which is why scaile writes those answers, fact-checks them and puts them to your team for approval rather than handing you a list of gaps.

The cited page belongs to somebody else. You are reachable, your page is good, and the assistant still quotes a comparison article, a directory or a competitor’s page. Then the work is partly off your domain: getting into the sources that get retrieved for that question, and publishing the comparison your buyers are reading elsewhere. Our case studies show what that shift looks like over a few months.

What can a manual check not tell you?

Three things, and it is worth being blunt about them.

Sample size. Five runs per question is a rounding error against the 60 to 100 runs the study itself put behind every prompt. Your fractions are directionally useful and statistically thin.

Market and language control. By hand you get one location, one language, one moment, so you cannot hold the market constant while varying the question. That is exactly what you need in order to know whether your German visibility differs from your Austrian visibility.

Change over time. A snapshot cannot tell you whether you moved. Rerunning by hand three months later gives you two snapshots taken with a shaky instrument, and the difference between them is mostly noise.

This is the honest reason the free AI Visibility Check exists. It reads your website, writes the ten buying questions your category actually produces, and puts every one of them to ChatGPT, Gemini and Perplexity from your market in your buyers’ language. You see the questions on the page and get the leaderboard and the cited sources by email. Beyond that first snapshot, tracking is a platform job rather than a spreadsheet job, because it means the same prompts run on a schedule with the movement and the cause reported together. That is not a thing a quarterly afternoon of manual prompting can approximate.

What is being named actually worth?

A shortlist position, and not much traffic. In Ahrefs’ March 2025 study of roughly 35,000 websites, run on data from July 2024 to January 2025, AI assistants accounted for about 0.1 percent of referral traffic, and Google sent 345 times more than ChatGPT, Perplexity and Gemini combined. That share has been climbing since, but it is still a rounding error next to search, and anyone selling AI visibility as a traffic channel is overselling it.

Attribution is genuinely hard on top of that. Google Analytics 4 now has an AI Assistants channel, but it only classifies a session when the medium is set to ai-assistant or the referrer matches a known list. Sessions that arrive without a referrer, and links a buyer copied out of an answer rather than clicking, fall into direct or unassigned. So your real AI-influenced traffic is larger than your dashboard shows and you cannot say by how much.

The value is upstream of the click. If an assistant names three suppliers and you are one of them, you are in the consideration set for a buyer who never opened a search results page. If it names three and you are not one of them, the deal was decided before anyone contacted you. That is worth measuring properly, which is a different claim from worth measuring with a stopwatch on the traffic report. If you are missing entirely, the common causes and the order to check them are the next page to read.

FAQ

Sources

• SparkToro and Gumshoe: AIs are highly inconsistent when recommending brands, 600 volunteers, 12 prompts, 2,961 runs, November to December 2025. Instability between runs for ChatGPT and Google’s AI, and the author’s suggestion to ask at least 60 to 100 times and average.
• Ahrefs: top mentioned sources are not shared across AI assistants, June 2025, ~76.7 million AI Overviews and ~950,000 prompts each in ChatGPT and Perplexity. Only 7 of the top 50 domains shared across all three.
• Ahrefs: only 12% of AI-cited URLs rank in Google’s top 10, 15,000 long-tail queries. Per-assistant overlap with Google’s top 10.
• Ahrefs: 38% of AI Overview citations pull from the top 10, March 2026, ~863,000 SERPs and ~4 million AI Overview URLs. The AI Overviews figure, down from around 76 percent in Ahrefs’ earlier count.
• Omniscient Digital: how LLMs source brand information, 240 prompts each containing a brand name, 23,387 citations, five platforms. 23 percent owned, 48 percent earned, 30 percent commercial content published by brands other than the one named.
• OpenAI: crawler overview. OAI-SearchBot, GPTBot, ChatGPT-User and OAI-AdsBot, and what each one is for.
• OpenAI: Memory FAQ. Memory and past chats personalise responses; Temporary Chat neither uses nor creates memories.
• Google Search Central: AI features and your website. Indexed and snippet-eligible as the only requirement, no special files or markup, query fan-out.
• Ahrefs: AI makes up 0.1% of traffic, March 2025, ~35,000 websites, data from July 2024 to January 2025. AI’s share of referral traffic, and Google against ChatGPT, Perplexity and Gemini combined.
• Google Analytics 4: default channel group definitions. The AI Assistants channel and the medium and referrer conditions it requires.
• Ahrefs on llms.txt adoption, 137,000 websites. 97 percent of the files found got zero requests in the month measured.