Who owns AI’s answers?

A weekly measurement of which pages AI assistants actually read before they recommend software.

By CiteGraph · Week of 2026-08-24 · 440 answers · 7 categories · 200 products observed

When a buyer asks an assistant for the best tool in a category, the model does not consult a ranking. It retrieves a handful of pages, reads them, and writes a recommendation from what it found. Which pages those are decides which products get named — and almost nobody knows which pages answer their own category. This report measures it, in the open, every week.

Why is there no single site AI assistants cite for software?

We expected concentration. The assumption behind most answer-engine advice is that a few review platforms — the G2s and Capterras of the world — are quietly deciding what models recommend, and that getting listed there is the whole game. That is not what the measurements show. Across 7 categories we found 12 distinct third-party domains supplying the pages behind leading recommendations, and not one of them appeared in a second category.

The gatekeepers are small. They are comparison posts on tooling blogs, engineering write-ups, community threads and, frequently, a rival’s own content marketing. Their authority is local to one question shape. Winning the page that answers “best AI video generator” teaches you nothing about the page that answers “best vector database”, because they share no publisher, no format and no audience.

Third-party domains that were the attributed source for at least one top-ten product, week of 2026-08-24.

Is being recommended by AI the same as being read?

No, and conflating the two is the most common mistake in this category. Two different events are worth counting. A mentionmeans the assistant named the product in its prose. A citation means one of the product’s own pages was used as a source for that answer. In our sample, 36 of 70 leading products had their own domain as the cited source. The rest are recommended constantly while being described entirely in someone else’s words.

This matters because the two numbers imply opposite actions. A high mention rate with low citation says the market already knows you and the model is repeating what others wrote — so the work is on those pages. A low mention rate with reasonable citation says your pages are being read but are not persuading — so the work is on your own words. Reporting a single “AI visibility score” collapses that distinction and tells a founder nothing about what to do on Monday.

Which products do AI assistants name most often?

Naming is concentrated: in an average category the top three products absorb about 54% of all naming among ranked products. But concentration at the top does not mean stability underneath, and it does not mean those leaders control their own story — the self-sourced column below is the one worth reading.

“Self-sourced” counts how many of a category’s top ten had their own domain as the cited page.

Category by category: who answers, and who is answered for

Each board below is public and updated weekly, with the questions asked printed in full and a 95% interval on every rate. The summaries here are drawn from the same data.

What this means if you sell software

The practical consequence is that answer-engine work is not a content-volume problem. There is a specific, findable page behind each question you lose, and it is usually one page rather than a genre of pages. Three things follow.

Definitions used in this report

How this was measured

For each category we write ten questions the way buyers type them — a straight pick, a comparison, a cheapest-option, an alternative-to-a-named-rival, and specific use cases — and put every one to three production AI assistants through their official grounded-search APIs, multiple times each. Nothing is scraped from a chat interface and no answer is simulated.

Products are extracted from the answer text itself rather than matched against a curated list, then resolved so that plans, tiers and renamed products count once. Cited pages are attributed to the product an answer attaches them to, not merely to the answer as a whole. Rates are sampled, so each carries a Wilson 95% interval on its board, and movement between weeks is only reported when it exceeds the two intervals combined.

Sample for this edition: 440 answers across 7 categories, 200 products observed, 81 ranked. Full method: citegraph.app/methodology.

Limitations

Answer engines change without notice, and so can these numbers. A sample of this size separates a real move from noise but cannot resolve two products a point apart. Questions are English-only and not region-specific. Category selection is ours, and reflects software categories we could ask credible buyer questions about — it is not a census of the software market. Attribution of a page to a product is inferred from the answer’s own citation structure, which engines format inconsistently. Where a number is uncertain, the interval is printed rather than rounded away.

Questions about this research