How Often Does AI Recommend You? Measure It, Then Raise It

AI engines answer the buyer's question and name the brands. When Google shows an AI Overview, the top organic result gets 58% fewer clicks, per an Ahrefs study of 300,000 keywords. Miss the list in the answer and the buyer never reaches your site.

Your analytics barely see it. AI chatbots sent 0.25% of referrals worldwide in August 2026. The choice happens inside the answer, before any click.

Most teams test by typing one question into ChatGPT. Ask again and the list changes. SparkToro found that ChatGPT and Google's AI repeat the same brand list less than 1 time in 100. One check tells you nothing. Hundreds of runs tell you where you stand.

1. Build the question set. We agree on 20 or 50 questions your buyers ask. They come from your sales calls, support tickets, and search data.

2. Sample the engines. I ask each question 10 or 20 times on each engine through its official API, with web search on. Every answer and every citation goes into a log.

3. Score with error bars. You get your share of answers and your citation share on each engine, next to three competitors. Every number carries a 95% interval.

4. Find the cause. Most AI crawlers do not run JavaScript, so I compare what they see with what your visitors see. I check which sites the engines cite in your place, and what those pages have that yours lack.

5. Fix and rerun. You get a ranked fix list. In the sprint we ship the fixes together, then rerun the same questions and test the change.

• A scorecard for each engine: share of answers and citation share, you against three competitors, with 95% intervals

• A citation map of the domains AI engines cite most for your questions, and where you stand on each

• A crawler report: robots.txt rules for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, and Google-Extended, plus a raw-HTML check of each key page

• A page review that scores your top pages for what engines quote: statistics, named sources, quotations, and direct answers

• A ranked fix plan with the effort and the owner for each item

• The question set, every raw answer as a CSV, and the Python scripts to rerun the audit

• A recorded readout call

B2B SaaS and devtool teams. Your buyers ask ChatGPT which tool to pick. You need to know whether you make the list, against whom, and on which engine. You want a number your CMO trusts, not a screenshot.

Maven instructors and course creators. Students ask Claude and ChatGPT which course to take. Your course page competes with Coursera, YouTube, and other instructors for that answer. The Snapshot shows where you stand for $950.

SEO and content agencies. Your client wants proof the GEO work paid off. I run the before-and-after measurement as an independent third party. You keep the client.

I measure what language models do. I hold a PhD in the Physics of Complex Systems and held a tenured faculty position at Aix-Marseille Université. I was a Data Science fellow at NYU. On Maven I teach LLM evaluation: paired tests, error bars, and eval gates. My Data For Science newsletter reaches 4,000+ readers. This audit points the same statistics at AI search.

How is this different from an AI visibility tool?

A tool shows a score. This audit shows the score, how far to trust it, why it sits there, and what moves it. Keep your tool. The audit tells you which of its swings are real.

Do you query the same ChatGPT my buyers use?

I sample through each engine's official API with web search on. API answers and app answers differ: the apps add memory, location, and account history. I spot-check 10 questions by hand in each app and report any gap.

Can you guarantee more AI mentions?

No. Nobody controls these engines. You get a measured baseline, fixes grounded in published research, and a test that shows whether they worked. In the KDD 2024 GEO study, adding statistics, quotations, and cited sources raised visibility by up to 40%.

What do you need from us?

Your domain, three competitors, and 20 questions your buyers ask. Read access to Google Search Console helps. Thirty days of server logs sharpen the crawler report. Neither is required.

Do you need access to our CMS or code?

Not for the audits. The sprint needs one contact on your web team, plus edit access or a pull-request workflow.

Which engines do you cover?

ChatGPT, Claude, Gemini, and Perplexity in every tier. The Full Audit and the sprint add Google AI Overviews. Ask for Copilot or Meta AI and I scope it.

Can we rerun the audit ourselves?

Yes. The Full Audit ships the question set and the Python scripts. You need an API key for each engine. The readout gives the API cost of one full rerun.

How long until fixes show up in AI answers?

Each engine recrawls on its own schedule. Your server logs show when each crawler comes back. The sprint waits four weeks after the last fix before the rerun.

Is our data safe?

Questions and answers go straight to the engines' APIs, never through a third-party tracking tool. An NDA is fine: send yours or ask for mine.

Can my company pay by invoice?

Yes. Maven issues company invoices.

What if the Snapshot says we are fine?

Then you have a baseline with error bars and a short list of risks. The $950 counts toward the Full Audit for 30 days.