The Pages Your Agency Built to Win AI Citations Just Lost Half of Them

OpenAI moved Free and Go accounts onto GPT-5.6 Luna on August 6 and retuned the model serving Plus and Pro the same day, adding a slider that lets paid users choose how much effort ChatGPT puts into a response. Peec AI ran its citation numbers that day. Listicles fell from 15.77% of ChatGPT's citations to 7.80%, a relative drop of 50.5%. Comparison pages fell from 9.08% to 6.17%, down 32.1%.

The fan-out queries went with them. Those are the searches ChatGPT fires off behind the scenes before it writes an answer, and the ones containing "best," "top," "vs," and "comparison" now make up a smaller share of the total. Fewer queries of that shape means fewer chances for a page built around that shape to get pulled in.

Product pages picked up the slack. They now account for 16.39% of retrieved pages, which puts them ahead of listicles. Asked about a category, the model has become more likely to go to the manufacturer for specifications and pricing than to a roundup written by a site earning commission on the products inside it.

ChatGPT is also doing more work per answer and crediting fewer sources for it. Prompts resolved with a single fan-out query dropped from 94.0% to 43.5%, and the average number of retrieved sources roughly doubled, from about twelve to about twenty-four. Olivier de Segonzac and the team at Resoneo, working separately, found unique domains cited per response falling from 19 to 15 after an earlier update, which points the same way from the other end.

Google moved twelve days later

Google started rolling out its third spam update of the year on August 18 and finished on the 21st, two days and sixteen hours later, across every language and location. The company called it routine, announced no new policies, and published nothing about it beyond a line in its ranking updates history.

What sits behind it happened in May. On the 15th, Google rewrote the opening line of its search spam policies so that the definition now covers attempts to manipulate generative AI responses in Google Search. Scaled content abuse, which describes pages produced in volume mainly to move rankings and names generative AI tools among the ways it happens, has been policy since 2024. The May edit stretched that policy from ranking to getting cited, and the August update is the first serious enforcement pass since it landed.

Blue-link ranking and AI Overview citation run on separate retrieval systems, and nobody is suggesting the two companies coordinated. Two systems that share nothing operationally moved against the same content format inside a fortnight. That says something about how much of it there is and how visible it had become to both of them.

What the retainer bought

AI visibility work sold since 2024 has been consistent in what it produces. Category roundups. "Best [category] software 2026." Head-to-head comparison pages, often naming competitors the company would never mention anywhere else. Alternatives pages. Published on a schedule, at volume, usually by an agency, usually on a line item labeled GEO or AEO.

Those pages worked, which is why so many of them exist. They were cheap to produce using the same models now declining to cite them, they were the highest-leverage format available at the time, and for roughly eighteen months they returned. Nobody bought them out of foolishness.

They are losing ground on both surfaces at once, and the reason is structural. A page that exists only to catch a comparison query has no audience underneath it. When retrieval behavior shifts, nothing holds it up.

What to do about it this week

Pull your own citation data and split it by page type, comparing July against the back half of August. Most AI visibility tools report aggregate mentions, and aggregate mentions cannot show you this.

Work out what share of everything you published in the last year was roundup or comparison format, and how much of that came from a vendor. Call the vendor and ask what they are telling clients about the August data. The quality of that answer is worth more than the deck they sent you in January.

Then look at your product and pricing pages, because that is where the retrieval is going. We have written before that most of what ChatGPT cites does not rank on Google. The Peec numbers add something to it. The model wants primary sources, in plain HTML, on the vendor's own domain, with real specifications and real prices on the page. B2B has spent ten years moving exactly that information behind a form.

The cheap version of this had a good run. What replaces it will cost more, because a page that earns a citation on the merits takes longer to write than a page shaped like the query it was built to catch.