How we track AI Overviews and AI citations for our own site

TL;DR: Every Monday morning we pull the same numbers: Search Console impressions, a fixed list of 30 questions checked against Google's AI Overviews, and the same questions asked to ChatGPT, Perplexity and Gemini. For each answer we record two things separately, whether we were mentioned and whether our site was cited. On day one the answer to both was zero, everywhere. That zero is the point of starting now.

Most advice about AI Overviews starts with a screenshot. Someone searches a question, sees their site in the AI answer, and writes a thread about how they did it.

Then they search again an hour later and it's gone. The thread stays up.

We wanted something duller and more useful: the same questions, asked the same way, every week, written down.

Where we started

When we set this up on 11 October 2026, nobody was finding rafalai.com through search or AI. Over the previous three months, Search Console showed 6 impressions and 0 clicks for queries, every one of them our brand name or a misspelling of it. Google had indexed 3 of our pages.

We then checked 12 of our target questions against Google. All 12 showed an AI Overview. We were cited in 0 of 12, and ranked in the normal results for 0 of 12. We asked 8 questions to ChatGPT, Perplexity and Gemini, 24 answers in all. We were mentioned or cited in 0 of 24.

That is a clean baseline. Anything that moves from here, we can see.

Google first, because Google reports it

For Google's AI features, the best source is Google itself. In June 2026 Google launched generative AI performance reports in Search Console, rolled out to all sites by 31 August. They show how often our URLs appeared in AI Overviews and AI Mode, broken down by page, country and device.

So the weekly pull starts there, along with the normal performance numbers: impressions, clicks and average position per page and per query. We read the page-level totals first. Search Console hides rare queries for privacy, so at our size the page totals are bigger and more honest than the query list.

We also check whether each page is indexed at all. Google's guide to generative AI features is plain that a page must be indexed and eligible to show with a snippet before it can appear in those features. For a new site, indexing is the first wall, and the clock on everything else starts when a page gets past it.

Then the questions, one by one

Search Console tells us where we appeared. It doesn't tell us what the AI Overview said, or who it cited instead. For that we keep a fixed list of 30 questions people ask about AI workflows and AI search, things like "what is an ai workflow" and "how to track ai overviews".

Each week a script runs every question through Google with the AI Overview loaded, using a search data provider, DataForSEO, from a fixed US location. For each one it records whether an AI Overview appeared and whether rafalai.com is in its sources.

The same questions go to ChatGPT and Perplexity every week, and to Gemini every other week. For each answer we store mentioned (yes or no), cited (yes or no) and which URL was cited.

Mentions and citations, kept apart

Being named and being linked are different results, and they fail in different ways.

An AI can know a business, name it, and link a directory instead. Or it can read a page, use its point, and credit someone else. If you count only mentions, you miss the missing link. If you count only links, you miss the business that is known but not credited.

So every row in our log has both columns.

Keeping the noise out

AI answers change from run to run. A citation that shows up once and disappears tells you very little.

We use two rules. A question counts as "cited that week" if any run cited us. And we only call a citation "held" once it appears two weeks in a row. That is slower than a screenshot, and far less exciting. It also means we won't announce a win that was really a coin landing on its edge.

We report counts with their denominator, "cited in 2 of 30", never a percentage built on a handful of answers.

Other checks in the routine

Some things decide whether an AI can cite us at all, so they're in the routine too.

Crawler access is one. OpenAI's crawler documentation says sites that block OAI-SearchBot "will not be shown in ChatGPT search answers," and recommends allowing it in robots.txt. We check that our pages return normally to the crawlers that matter, and that our robots.txt and our network settings say the same thing.

We also keep an llms.txt file, following the llms.txt proposal, and update it when we publish. Google says its Search ignores the file, so we don't expect it to move our Google numbers. We keep it because agents read it, and because it costs us almost nothing to maintain.

The weekly log and readout

The numbers go into one spreadsheet: date, page, question, source, metric, value. Raw results are saved by date, so we can recheck anything later.

Then a short readout: what moved, which pages got indexed, anything broken. If nothing moved, the readout says so in one line. Most weeks, for a while, that is what it will say.

We note each article's format, answer-first or normal, in the same log.

If you want to copy it

You don't need our script. You need the habit: Search Console's generative AI report once a week, a fixed list of questions, the same tools each time, two columns for mentioned and cited, and counts instead of percentages.

We'll publish what the log shows as it fills in, including the weeks where the answer is still zero.