Anything Engine Optimization The rolling record of the AI-search industry.
A single summation sign on a deep colour field — the house card for the weekly briefing.
AEO Wire
AEO Wire · Sep 28 AEOGoogle SearchEngines

Last Week in AI Search: September 28

What this answers

How many items did the AI-search wire publish this week?
32 items between 2026-09-21 and 2026-09-27, across the lanes AEO, Google Search, Engines. Each is dated and sourced on the wire.

In this issue: Atlantic robots.txt allows GPTBot, Cloudflare blocks it · Study: Google AI Mode cuts clicks and erodes user trust · Google tests AI Overviews below hotel…

The Story

Scrunch analyzed about 10,000 pages that AI assistants had already cited between May 17 and June 30, across ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, AI Overviews and AI Mode. Roughly 30% of the comparison pages were a company ranking its own product first.

When an answer cited one of those lists, it named the list’s author in about 39% of answers, against 19% when it didn’t. It recommended the author about 7% of the time, against 4%. In 63.9% of those answers, it recommended no one.

The smaller number is the sturdier one. The 4%-to-7% recommendation lift survived Scrunch’s test for brands that were already rising; the larger mention lift hasn’t been through that test yet.

And there is less room to be named at all. In a fixed panel of B2B tech prompts, ChatGPT named 3.87 brands per answer in September, down from 5.36 in March, per BrightEdge. Transactional answers moved the other way, which BrightEdge left out of its headline.

What the record says to do

  • Getting cited comes first. Scrunch measured only lists an assistant had already cited, so it says nothing about lists that weren’t. For Google ranking, an old or unreadable date cut a listicle’s odds of the top three by 56%, per Growth Memo. Keep dates visible and machine-readable and links crawlable.
  • Size it for the outcome. For AI recommendation, lists of four to 10 brands did best in Scrunch’s data. For Google ranking, titles promising 51 to 100 items ranked higher than ones promising two to five, per Growth Memo. Different outcomes, different lengths.
  • Treat reputation as the bigger bet. For whether a model looks you up at all, models searched for brands they already knew 3.2 times as often, per geoSurge. For recommendation, in a controlled test, a rival needed less than a 0.1-star rating edge to overturn a leader. The evidence points to reputation as the larger long-term bet, not pages you write about yourself.

Scrunch and BrightEdge are vendors publishing research tied to their products. Read the item → · where this sits: I’m getting cited. Am I getting recommended?

The Follow-Up

Last week’s Story was crawler access, and the record moved in both directions. The Atlantic now allows GPTBot in robots.txt while still blocking it at the homepage. El País let ClaudeBot back in even though its robots.txt still bars it. The policy file and the live server can now disagree, and reverse, within days. The standing answer changed to say so.

The Numbers

  • ChatGPT-5.6 Luna, now the default model, used a site: search in 76% of its fan-out queries, up from 21.6% for GPT-5.5, and consulted 43% fewer sources. (Seer Interactive; 346 matched prompts) Read →
  • Adthena’s observed weekly count of ChatGPT advertisers rose 65% in eight weeks, from 7,306 to 12,075 between July 9 and August 31, before ChatGPT Ads passed 60 countries in September. (Adthena study) Read →
  • ChatGPT’s share of AI assistant prompts in Comscore’s panel fell from 70% to 50% between January and June, while Gemini’s rose from 17% to 30%. (Comscore panel) Read →
  • ChatGPT still sent 95.1% of AI referral traffic to websites in August, a 2026 high. Losing prompt share and keeping the clicks are different markets. (BrightEdge; method undisclosed) Read →

The Rest

  • A second randomized study, after the first one in last week’s issue, found that forcing AI Mode on searchers cut clicks to outside sites and lowered their trust. Read →
  • Google is opting eligible sellers, Shopify stores among them, into checkout inside AI Mode and Gemini by default; keeping sales on your own site means switching it off. Read →
  • Search Console now splits “Web” results into text-based and multimodal search. (spotted in Google’s documentation) Read →
  • Spammers are publishing public Gemini Notebook pages to ride the product’s domain authority into Google’s index. Read →
  • Google is testing AI Overviews below the hotel knowledge panel. (single sighting; not confirmed by Google) Read →
  • Google is testing AI Overview links that open AI Mode instead of a website. (single sighting; not confirmed by Google) Read →
  • Ask YouTube now sorts product-review videos into comparison tables. Read →

All 32 items from September 21–27, dated and sourced: the week on the wire · State of Search · Corrections · Archive. Seen something the record should hold? Reply to this email; tips are credited by name.

The full wire

32 items published September 21–27. Every item dated, sourced and linked. An item can carry more than one lane, so lane totals exceed the count.

Google Search · 11

Engines · 9

Research · 7

Tools & Vendors · 4

Law & Money · 1

Read the complete archive →

Why it matters: The one place on the wire where the week is read rather than reported — compiled from the published record, framed by the editor, checkable line by line.

Glossary: Answer engineAI visibilityAI OverviewsAI ModeQuery fan-out

Via AEO Wire ↗

Posted to the wire September 28, 2026. Updated September 29, 2026. Edited by Joe Balewski.