How to Track AI Search Visibility When the Platforms Give You Nothing

You can track AI search visibility, but not the way you track Google rankings. There is no console, no impression count, no query report and no position history from any of the assistants. What you can do is sample: run a fixed set of buyer questions through ChatGPT, Perplexity, Gemini, Claude and Google AI Mode on a schedule, record whether you were named or linked, count AI crawler hits in your server logs, and isolate the referral traffic that does carry an assistant referrer. That gives you a rate and a trend rather than a number. A rate and a trend is enough to run a programme on. A vanity score is not.

If you have been handed a dashboard showing "AI Visibility: 34" and asked whether that is good, this post is the answer. The honest version is less satisfying than the dashboard and considerably more useful.

There is no Search Console for ChatGPT

AI search visibility is the degree to which large language model assistants mention, recommend or link to your brand when someone asks them a question your business could answer. It is not a metric any platform reports to you. Google Search Console tells you impressions and average position because Google chose to build that product. OpenAI, Anthropic and Perplexity have built nothing equivalent, and there is no signal in their public product surfaces that they intend to.

Every number you will ever see about your AI visibility comes from one of three places: your own analytics, your own server logs, or somebody running prompts and reading the answers. There is no fourth source. A vendor score is the third of those aggregated, and its quality depends entirely on which prompts were run, how often, from which accounts and on which model version. Those choices are usually undisclosed.

The useful reframe: you are not doing analytics, you are doing market research. Researchers do not get a census either. They build a repeatable instrument, run it on the same population on a fixed schedule, and read the change. That practice is cheap, honest about its error bars, and it will tell you within a quarter whether your work to get cited by ChatGPT and Perplexity is moving anything.

The four things actually worth measuring

Most AI visibility programmes track one thing, citations, and ignore the other three. That is a mistake, because citations are the noisiest of the four and the least connected to revenue.

The fourth row is the one almost nobody tracks and the one that most often produces an immediate, fixable problem. An assistant that confidently tells a buyer you only serve enterprise clients, or that you sell a product you have never offered, costs you deals in a way no citation count will reveal. It is also the cheapest item to fix, because the usual cause is stale or contradictory information on pages you control.

Measuring accuracy takes two prompts run against every assistant on your list, monthly: "What does [company] do?" and "What are the risks of working with [company]?" The second one is uncomfortable and it is the most informative prompt in the panel, because it surfaces whatever negative or confused framing the model has absorbed.

Code each answer against a short fact sheet: what you sell, who you serve, where you are based, what you explicitly do not do. Then count the error rate. An assistant getting two of six facts wrong is a measurable defect with a traceable cause.

The left column is a two-week job. The right is a quarter or more of outreach and patient republishing. Knowing which column an error sits in is the whole value of measuring accuracy.

How do you find AI assistant referral traffic in analytics?

Start with the referrer report and build a channel group for assistant domains. The list moves, so audit it quarterly rather than setting it once. As of late 2026 the domains worth watching are chatgpt.com and chat.openai.com, perplexity.ai and its www subdomain, claude.ai, gemini.google.com, copilot.microsoft.com, and you.com. In GA4 you would create a custom channel group matching session source against those hostnames. In a server-side or log-based analytics stack you match the Referer header directly, which is cleaner.

Now the caveat that most agencies skip. A significant share of assistant-influenced traffic carries no referrer at all. It arrives as direct. This happens for several independent reasons: some assistant surfaces open links in a way that strips the referrer, some users read the answer and then type your name into a browser bar later, some are in a desktop or mobile app rather than a browser, and referrer policies on the originating page can suppress it. You cannot distinguish that traffic from someone typing your URL from memory.

So your assistant referral number is a floor, not a total, and you should report it as one. Watch direct traffic alongside it, specifically direct traffic to deep pages. Nobody types a URL like /services/erp-development from memory. A sustained rise in direct sessions landing on interior pages, with no campaign or email send to explain it, is the common fingerprint of assistant-driven discovery. The same logic applies to Google AI Mode, where the optimisation and measurement rules differ again.

• Build the channel group once, audit it quarterly. New assistant surfaces appear and old hostnames get retired, so a channel group set in January will silently miss traffic by June.
• Segment direct traffic by landing page depth. Direct sessions to the homepage are ambiguous. Direct sessions to a specific service or blog URL are almost never typed from memory.
• Watch branded search volume in Search Console in parallel. Assistants that recommend you without linking still generate name searches, which shows up as branded query growth with no ad spend behind it.
• Keep the raw referrer values. Aggregated channel reports lose the subdomain detail you need to work out which surface actually sent the visit. And you cannot UTM-tag a link you do not control, so ignore any vendor promising per-assistant tagging.

Server logs are your only first-party signal

Everything else in this post is either sampled or incomplete. Server logs are neither. When an AI crawler fetches a page, that request lands in your access log with a timestamp, a URL, a status code and a user agent string. You own that record completely.

The crawlers worth isolating are named and documented. GPTBot handles OpenAI training crawls, OAI-SearchBot handles search indexing for ChatGPT, and ChatGPT-User fires when a user's live request causes a fetch. PerplexityBot crawls and Perplexity-User fetches on demand. Anthropic operates ClaudeBot along with user-triggered fetchers. Bingbot matters because Microsoft Copilot leans on the Bing index. Applebot and Amazonbot are worth logging even if you are not optimising for them yet. Google-Extended is different and people get this wrong constantly: it is a robots.txt control token for Gemini and Vertex AI training, not a separate crawler with its own user agent, so you will not see it in your logs at all.

What the logs give you is coverage: which of your pages the assistants have actually seen, how recently, and how often they come back. A page that has never been fetched by OAI-SearchBot cannot be cited by ChatGPT search. That is a hard, checkable fact, and it turns a vague complaint ("we are not showing up in AI answers") into a specific to-do list ("these eleven service pages have zero AI crawler hits in ninety days"). Crawl frequency is also a decent proxy for perceived importance, and it is the fastest way to confirm that a schema markup rollout aimed at AI citations is at least reaching the machines it was written for.

Build a fixed prompt panel of 30 to 50 buyer questions

A prompt panel is a fixed, written-down list of questions that real buyers ask, run through the same set of assistants on the same schedule so the results are comparable over time. It is the closest thing to a rank tracker that AI search allows, and the discipline that makes it work is the word "fixed". The moment you start changing the prompts between runs, you lose the ability to say anything about the trend.

Three runs across five assistants on forty prompts is six hundred observations. That sounds heavy until you do it, at which point it is roughly a day of focused work per month, and less once you have a template. Doing it by hand for the first two quarters teaches you things about how the assistants reason that no scraped dataset will.

The reason this works and single-answer screenshots do not is variance. Ask the same assistant the same question twice and you will often get different sources, different ordering and sometimes a different recommendation. Personalisation, retrieval freshness, provider-side A/B tests and the account you are logged into all move the output. Anyone sending you a screenshot as proof of AI visibility, good or bad, is showing you one draw from a distribution. You track rates across a panel. You never treat a single answer as data.

What AI visibility tools can and cannot do

A category of tooling now automates the prompt panel. It splits into three types: prompt-monitoring platforms that run your questions on a schedule and chart mention rates, log-analysis tools that isolate AI crawler activity, and modules bolted onto existing SEO suites that blend both. I am deliberately not naming products with accuracy claims attached, because nobody, vendors included, can validate an accuracy figure against ground truth that does not exist.

What automation genuinely buys you is sample size and consistency. A tool running your forty prompts twenty times a week produces a far smoother signal than you running them three times a month, and it will not quietly drift the wording the way a human does. That is real value and worth paying for once your panel has stabilised.

Buy tooling for throughput, never for truth. Your panel is the instrument; the tool is a faster way to run it. If you cannot explain what a number means without opening the vendor's documentation, keep it out of the report. The same scepticism applies to AI-SEO advice generally, which is why we wrote that llms.txt is not the requirement it is sold as.

A measurement cadence you could start on Monday

Here is the actual operating rhythm. It fits in roughly half a day a week for one person and produces something defensible at the end of a quarter.

That last row is where most AI visibility programmes fall apart. Citations are not an outcome. A citation on a question no buyer asks before spending money is worth nothing, while a mention on the comparison question your sales team hears every second call is worth a great deal. Weight the panel by pipeline value: tag each prompt with its deal stage, and report shortlist-stage movement separately from category-education movement.

Then close the loop on the sales side. Add one question to your intake form or discovery call: "how did you first come across us?" with an explicit option for "an AI assistant recommended you". Self-reported attribution is imperfect, but it is the only bridge between a mention rate and a closed deal, and over a few dozen conversations it corroborates or contradicts your panel better than any model does. If your panel shows you invisible on the questions that matter, the fix is not more measurement, it is the underlying work that structured organic search programmes have always required.

Common questions

Is there any official way to see how often ChatGPT cites my website? No. OpenAI does not publish a reporting surface equivalent to Google Search Console, and neither does Anthropic, Perplexity or Google for its AI Mode answers. The only observable signals you own are server log entries from crawlers such as GPTBot and OAI-SearchBot, and referral sessions in your analytics from chatgpt.com. Everything else comes from sampling prompts yourself.

Why does my AI referral traffic look so small compared to the mentions I see? Because referral tracking undercounts by design. Assistant answers often lead to no click at all, and when a click does happen the referrer is frequently stripped by app surfaces, referrer policies or the user simply searching your brand name later. Treat the referral figure as a floor, watch direct sessions to deep interior pages alongside it, and monitor branded search growth as a third corroborating signal.

How many prompts should an AI visibility panel contain? Thirty to fifty is the working range for most businesses. Fewer than thirty and month-to-month noise swamps the signal, because answers vary between identical runs. More than fifty and the manual effort becomes hard to sustain, which leads to skipped months and a broken trend line. Split the panel across category, comparison and brand questions, and keep the wording frozen between runs.

Can I trust a vendor AI visibility score? Only if they disclose the methodology. Every score in the market is derived from sampled prompts, so its value depends on which prompts were run, how many times, against which model versions, from which country and account state. Ask for those details in writing. A vendor who will not share the prompt list is showing you a number with no denominator behind it.

What does it mean if an AI assistant describes my business incorrectly? It means a buyer researching you is being given wrong information at the exact moment they are forming an opinion. The cause is usually contradictory or stale descriptions on your own pages, missing Organization schema, or an outdated third-party profile the model trusts. On-site causes are fixable within a sprint. Off-site ones take a quarter or more of corrections and republishing.

Do server logs really help if the citation itself is invisible? Yes, because they answer a prior question. If OAI-SearchBot has never fetched a page, ChatGPT search cannot cite it, so zero crawler hits is a definitive diagnosis rather than a guess. Logs also show crawl frequency and recency per URL, which tells you whether new content is being picked up and how fast. Verify crawler identity by reverse DNS, since user agent strings are easily spoofed.