Are AI Engines Recommending Your Business? How to Measure Your AI Search Visibility
Short answer: yes, you can measure whether AI engines are recommending your business, but not with a single number and not from your analytics dashboard alone. The reliable approach is to triangulate four methods: manual prompt testing across the major engines, a dedicated AI visibility tool, server log crawler analysis, and a GA4 referral channel. Each one sees a different slice, and none of them is complete on its own, because most of your presence inside AI answers never produces a click you can count. This guide walks through all four, plus the traps that make people badly misread their own data.
If you want to measure AI search visibility honestly, the first thing to accept is that the old habit of typing your brand into Google and eyeballing the results tells you almost nothing anymore.
Why "just check Google" stopped telling you anything
The surface where buying decisions get shaped has moved inside AI answers, and that surface is enormous. Google's AI Overviews reached two billion monthly users by mid 2025, its conversational AI Mode passed one billion monthly users at Google I/O in May 2026, and ChatGPT reached 800 million weekly active users by late 2025 and has kept growing. Your customers are asking these systems the questions they used to type into a search box.
The catch is that answering the question in place removes the click. Pew Research Center found that when an AI summary appears, users click a traditional search result in only 8 percent of searches, versus 15 percent when there is no summary, and they click a link inside the summary itself just 1 percent of the time. Zero click behavior is now the norm rather than the exception: Similarweb clickstream data put roughly 68 percent of Google searches ending without a click in 2026, approaching 80 percent on desktop for business audiences. This is the same shift we described in The End of the Ten Blue Links and, for local operators specifically, in The Visibility Squeeze. If your only measurement instrument is click traffic, you are trying to measure an iceberg by the tip.
First, get precise: mentioned, cited, and recommended are three different wins
The most common measurement mistake is collapsing everything into one "are we visible" number. Being present in an AI answer has three distinct levels, and they carry very different value:
- Mentioned. The engine names your business somewhere in its answer.
- Cited. The engine links to your own domain as a source for its answer.
- Recommended. The engine puts you first, or names you as the pick, when a user asks who is best.
A brand can be mentioned constantly and recommended never. Tracking only "mentions" will flatter you; tracking only "am I the top pick" will scare you. Measure all three separately. The mechanics of how engines decide who gets cited, and how you influence that, are the subject of AI Citation Mechanics. Measurement is how you find out whether any of it is working.
Four ways to measure AI search visibility
No single method captures the whole picture, so treat these as four instruments pointed at the same object from different angles. The table summarizes what each one actually sees.
| Method | What it shows | Cost | Main blind spot |
|---|---|---|---|
| Manual prompt testing | Whether you are mentioned, cited, or recommended for real buyer questions | Free (your time) | Labor intensive and noisy; hard to scale |
| AI visibility tools | Share of voice, citations, and sentiment tracked over time at scale | Subscription | Costs money; each vendor samples differently |
| Server log crawler analysis | Which AI systems are fetching your pages at all | Free (your logs) | A crawl proves fetching, not that you were cited |
| GA4 referral tracking | Users who clicked through from an AI engine to your site | Free | Misses the majority of visibility that sends no click |
1. Manual prompt testing, the free baseline
Start here because it costs nothing and it is the ground truth every tool is trying to approximate. Build a fixed set of 20 to 50 questions your customers actually ask, phrased naturally: "best [your category] in [your city]", "alternatives to [a competitor]", "is [your brand] any good". Run that same set across the engines that matter to your audience, which today usually means ChatGPT Search, Google AI Mode and AI Overviews, Perplexity, Gemini, and Microsoft Copilot. For each answer, record the three outcomes above (mentioned, cited, recommended) and count your appearances against named competitors to get a share of voice.
The discipline that separates a real measurement from a vanity check is sampling. AI answers are not deterministic. When SparkToro ran the same brand recommendation prompts thousands of times, it found that any two answers returned the same brand list less than 1 percent of the time. So run each prompt ten or more times, in a logged out or incognito session to blunt personalization, and record frequency rather than a single yes or no. One check is not a measurement. We unpacked why every model returns a different world in The Shattered Mirror.
2. Dedicated AI visibility tools
A whole category of software now automates that prompt testing loop at scale, tracking your mentions, citations, share of voice, and sentiment across engines over time. These are the pure play options worth knowing. Pricing in this category moves quickly, so treat the figures below as an approximate July 2026 starting point, not a quote.
| Tool | Distinguishing focus | Approx. entry price |
|---|---|---|
| Am I Cited | Daily citation and sentiment monitoring across eight engines; free trial with no card | Free tier / trial |
| Rankscale | Credit based GEO tracking with a competitor benchmark and readiness audit | from ~$17/mo |
| Otterly.ai | Mentions, average position, and share of voice, plus on page GEO recommendations | from ~$29/mo |
| Peec AI | Visibility, position, and sentiment by model and region, with prompt tagging | Mid market subscription |
| Profound | Enterprise answer engine insights: how AI represents your brand in conversations | Enterprise |
The large SEO platforms have moved into this space as well, so if you already pay for one, check whether your plan includes AI visibility tracking before buying a second tool. Semrush and Ahrefs both shipped AI visibility features in the last half year that draw on very large prompt databases. Whichever you choose, the buying question is simple: does it track the specific engines your customers use, and does it separate being mentioned from being cited from being recommended?
3. Server log crawler analysis
Before an engine can cite you, it has to fetch your pages, and that leaves a fingerprint in your server logs. Grep your access logs for the user agent tokens of the AI crawlers. The token names are stable even though their version suffixes increment over time, so match on the name.
| Operator | User agent tokens | What it means |
|---|---|---|
| OpenAI | GPTBot, OAI-SearchBot, ChatGPT-User | Training crawl, search index crawl, and live user triggered fetch |
| Perplexity | PerplexityBot, Perplexity-User | Search index crawl and live user triggered fetch |
| Anthropic | ClaudeBot, Claude-SearchBot, Claude-User | Training crawl, search quality crawl, and user initiated fetch |
| Google-Extended | A robots.txt control token for Gemini, not a separate crawler; fetching is still done by Googlebot |
The token names and behaviors above come straight from the operators' own docs for OpenAI, Perplexity, Anthropic, and Google. Reading logs proves an engine is interested in your pages, which is a necessary precondition to being cited. It does not prove you were cited, and the gap between the two is vast. Cloudflare measured the crawl to click ratio across its network and found that in mid 2025, for every visitor an engine referred back, Google crawled about 5 pages, OpenAI about 1,091 pages, and Anthropic about 38,065 pages. AI systems consume far more than they send back, which is the clearest possible proof that click traffic wildly understates real visibility.
4. GA4 referral tracking
The users who do click through from an AI engine show up in your analytics, and you can isolate them. In GA4, create a custom channel group (Admin, then Data Display, then Channel Groups) called something like AI Traffic, defined as sessions whose source matches a regular expression of AI hostnames. A workable list of referrers includes the following:
chatgpt.com | chat.openai.com | perplexity.ai | claude.ai | gemini.google.com | copilot.microsoft.com | bing.com/chat | deepseek.com | grok.com | meta.ai | you.comHere is the caveat that keeps this method honest: referral traffic is a floor, not a measure. Between 35 and 70 percent of AI referral sessions arrive with no referrer header at all, because the AI app's embedded browser strips it, so those visits land in your Direct bucket and never get attributed to AI at all. GA4 tells you who clicked. It cannot tell you who was recommended and did not click, and given the zero click data above, that is most of your visibility.
The traps that make people misread their own data
Every method above has a failure mode, and knowing them is the difference between a measurement and a delusion.
- Non-determinism. One prompt run once tells you nothing, for the reasons above. Sample and report frequency.
- Unstable and hallucinated citations. A citation you saw once may not be real or durable. A Columbia Journalism Review study of eight AI search engines across 1,600 tests found they gave incorrect citation answers more than 60 percent of the time, with some engines pointing over half their citations at fabricated or broken URLs. Verify that a citation actually resolves to your page. We wrote about why these after the fact citations are so slippery in The Bibliography Is Not the Brainstorm.
- Personalization. Your logged in view reflects your own history, not your customer's. Test logged out and, where you can, across locations.
- Prompt phrasing sensitivity. Small wording changes flip which brands appear, so freeze your prompt set and version it before you compare month to month.
- Crawled is not cited is not clicked. Each method proves a different thing. Only triangulating all of them gives you a defensible read.
One point of intellectual honesty: the headline Pew finding on click rates has been publicly disputed by Google, which questions the study's methodology. The broader direction is corroborated by independent clickstream data, but a good measurement practice acknowledges the debate rather than citing one study as settled fact.
A measurement routine you can start this week
You do not need all four instruments running on day one. A realistic starting cadence looks like this:
- Write down 20 to 30 buyer intent prompts and run them, logged out, across ChatGPT, Google AI Mode, Perplexity, and Gemini. Log mentioned, cited, and recommended for each, plus your share of voice against competitors.
- Add the AI Traffic channel in GA4 today; it takes ten minutes and starts collecting immediately.
- Once a month, grep your server logs for the crawler tokens above to confirm the engines are fetching your key pages.
- If you have budget, add one visibility tool to automate the prompt loop and watch the trend rather than a snapshot.
Run that quarterly at a minimum, monthly if AI referrals are already meaningful for you, and you will have a defensible picture of whether your presence in AI answers is rising or falling, which is far more than most of your competitors can say.
Measuring the problem is the first half. Fixing it, so that the engines mention, cite, and recommend you more often, is the harder half, and it is what we do every day. If you want help turning this into a repeatable system for your business, reach out to our team and we will walk you through it.
Citations
- Digiday "Google's AI Overviews reach over 2 billion monthly users" (July 24, 2025)
- Google "Google Search's I/O 2026 updates" (May 19, 2026)
- TechCrunch "Sam Altman says ChatGPT has hit 800M weekly active users" (October 6, 2025)
- Pew Research Center "Google users are less likely to click on links when an AI summary appears" (July 22, 2025)
- Similarweb "Zero-Click Marketing: What the 2026 Data Means" (June 10, 2026)
- SparkToro "AIs are highly inconsistent when recommending brands or products" (January 28, 2026)
- Cloudflare "The crawl-to-click gap: AI bots and the web" (August 29, 2025)
- Columbia Journalism Review, Tow Center "AI Search Has a Citation Problem" (March 6, 2025)
- OpenAI "Overview of OpenAI crawlers and user agents" (accessed July 14, 2026)
- Perplexity "Perplexity Crawlers" (accessed July 14, 2026)
- Anthropic "Does Anthropic crawl data from the web, and how can site owners block the crawler" (April 7, 2026)
- Google "Google's common crawlers" (accessed July 14, 2026)
