AI Search Analytics: The Numbers You Can Actually Prove
- The AI search numbers you can prove come from your own server: crawler reads per page, visits from an AI referrer or AI utm tag, and goals those visitors complete.
- Position in an AI answer, mention share of voice, and the wording of the answer itself are not in your logs. No tool can prove them from your site's data.
- A 600-person SparkToro study in January 2026 found fewer than 1 in 100 runs of the same prompt returned the same list of brands.
- Prompt-sampling tools report a number that shifts with server load. One lab got 80 unique completions from 1,000 runs of an identical prompt at temperature 0.
- Kymo samples no prompts. It records what AI engines did on your own pages, per page.
Most AI search analytics is inferred. The provable kind is narrower, and it is the part you can act on. You can prove that a specific crawler read a specific page, that a visitor arrived with an AI referrer or an AI utm tag, and that a visitor whose first touch was AI completed a goal. Everything else is someone's model of someone else's model.
A table of AI search numbers, and whether your data can prove them
| The number | What it actually measures | Provable from your own data? |
|---|---|---|
| Crawler reads per page | Which AI crawler fetched which URL, and when | Yes. It hits your server. |
| AI traffic | Visits tagged with an AI referrer or an AI utm source | Yes. |
| Goals from AI traffic | Completions by visitors whose first touch was AI | Yes. |
| Rank in an AI answer | Where your brand sat in one generated response | No. The response never reaches your logs. |
| Share of voice | Your mentions as a share of all mentions in sampled answers | No. It depends entirely on the prompts sampled. |
| What the assistant said | The wording of the answer | No. You would need the answer. |
The right column is the whole argument. Three rows are measurements. Three rows are estimates dressed as measurements.
Prompt sampling produces a number that changes on its own
A tool that types prompts into assistants and counts your brand is measuring a sample. The sample moves.
SparkToro, working with Gumshoe.ai, ran 12 prompts through ChatGPT, Claude and Google's AI Overviews and AI Mode 2,961 times with 600 volunteers in January 2026. Fewer than 1 in 100 runs returned the same list of brands. Fewer than 1 in 1,000 returned the same list in the same order. Visibility share across many runs held up better than position did. Gumshoe sells AI tracking, and the study discloses that.
The instability has a mechanical cause. Thinking Machines Lab ran one prompt 1,000 times at temperature 0 in September 2025 and got 80 unique completions. Server load changes the batch size, and the batch size changes the output. Temperature 0 is not a synonym for deterministic.
Rand Fishkin, who co-founded SparkToro, called tools that report a ranking position in AI "full of baloney." The measurement physics backs him up. You can compare that sampling approach against the rest of the market in The Best AI Visibility Tools, and What Each One Can't Tell You.
The model changes under you, so a snapshot has a shelf life
Even a perfect sample describes a moving target.
Chen, Zaharia and Zou found GPT-4 scored 84% on a prime-number task in March 2023 and 51% on the same task in June 2023. OpenAI rolled back a GPT-4o update on April 29, 2025 because the model had become overly flattering. A visibility score captured in January describes a model that may not exist in March.
The AI surfaces disagree with each other too. Ahrefs compared 540,000 query pairs in December 2025 and found Google's AI Mode and AI Overviews cited the same URLs only 13.7% of the time, with 86% semantic similarity between the queries. There is no single AI answer to rank in.
Mentions, position and citations are three different numbers
This is where scores get slippery. Digital Applied took one brand and one dataset in 2026 and scored it three ways: 20% on mention-based share of voice, 16.8% position-weighted and 31.4% citation-based.
Three methods, three answers, one dataset. When that happens, the score says more about the method than about the brand.
Clicks out of AI answers are measurable, and smaller than the old ones
Pew Research Center surveyed 900 US adults in July 2025. On Google pages with an AI summary, users clicked a result in 8% of visits. Without a summary, 15%. They clicked a link inside the summary in 1% of visits.
That is the squeeze. Fewer clicks out of search overall, and a slice of what remains passes through a summary that mostly does not link. It shows up in your own numbers long before it shows up in anyone's report.
What Kymo records, and what it refuses to guess
Kymo is a hosted analytics platform with one dashboard for human visitors and AI crawlers. It counts visitors without cookies and without localStorage. The visitor identifier is a salted server-side hash that rotates at UTC midnight, so unique visitors are unique per day. Raw IPs are never stored.
AI crawler tracking is a server-side receiver that classifies crawler user agents into three categories. ai_answer is a live fetch made because one person asked a question right now, like ChatGPT-User. indexing is a crawler building a search index for later answers, like Claude-SearchBot. training is a bulk crawl gathering model training data, like Bytespider.
Crawler reads do not count against your event limit. Solo is $9 a month, or $90 a year, with up to 10,000 events a month. Studio is $29 a month, or $290 a year, with up to 100,000 events a month. Every feature is on both plans. Both start with 14 days free, no card required. Go over your limit and the dashboard pauses. Your data is not deleted.
What Kymo does not do: it does not track mentions, it does not sample prompts, and it does not tell you what an assistant said. It cannot report your rank in an AI answer, because there is no stable rank to report. It reports a read, a visit and a sale, per page. For the wider picture, read AI visibility, explained.
Two things worth knowing before you build a plan. Publishing an llms.txt file is a cheap bet, not a ranking lever. Adoption is partial, no major operator has publicly committed to honouring it, and publishing the file is not the same as a crawler fetching it. Only your own server sees which one happened. Second, blocking a training crawler does not remove you from that operator's AI answers, because the indexing and live-fetch agents are separate tokens.
Start with Reading the AI report and keep Kymo documentation open in a tab. For the wider playbook, see AEO explained and The Small Business Guide to AEO.
Stop guessing. Watch.
Most AI visibility tools ask a chatbot and report a guess. Kymo watches your own site instead: which pages AI engines read, how many people their answers send you, and which of those people buy. Every number comes from your server, not from a prompt someone else chose.
Start free → 14-day free trial. No card required.