Anything Engine Optimization The rolling record of the AI-search industry.
Send a tip
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

This week in one minute. Paywall or block? The Atlantic’s robots.txt became readable again and now names GPTBot as allowed, but its live homepage still 403s GPTBot itself — stated policy and enforcement now disagree, the mirror of El País’s September 20 reversal. Costing traffic? A second randomized field experiment, this one testing AI Mode rather than AI Overviews, found a forced AI Mode condition cut click-through to external sites and also lowered searchers’ trust, satisfaction and sense of relevance — worst among heavy searchers — while confirming the other randomized study’s finding that removing AI Overviews raises clicks without hurting satisfaction. Where does AI now appear in Google Search? Google was seen testing the AI Overview on hotel search results moved below the hotel knowledge panel instead of above search results, in screenshots an SEO professional shared and Search Engine Roundtable reported. Buy your way in? Two new Adthena readings landed the same day.

The Monday map

Atlantic robots.txt allows GPTBot, Cloudflare blocks it

Study: Google AI Mode cuts clicks and erodes user trust

Google tests AI Overviews below hotel knowledge panels

ChatGPT advertisers jump 65% as AI Overview queries shorten

Is AI-crawler access now a paywall rather than a block?

Before: The Atlantic’s robots.txt had returned a 403 at every capture since September 13, and Search Engine Land’s robots.txt had been readable throughout the panel’s tracking.

Now: The Atlantic’s robots.txt became readable again and now names GPTBot as allowed, but its live homepage still 403s GPTBot itself — stated policy and enforcement now disagree, the mirror of El País’s September 20 reversal. Search Engine Land’s robots.txt started returning the same Cloudflare challenge as its homepage, closing off the last part of the site the panel could read.

The standing answer: On the Cloudflare-proxied publishers the wire tracks, yes for some — access is starting to split by whether the crawler’s operator pays rather than by whether the site allows AI crawling at all, and the split is now moving in both directions on the same sites within days.

Evidence: first-party data, Crawler access panel

Still open: Whether the GPTBot block at the four Cloudflare sites is a publisher decision naming OpenAI specifically or a Cloudflare-side default reclassifying GPTBot — the panel logs the HTTP response, not who set it or why.

Read the item →

Are AI answers costing me traffic?

Now: A second randomized field experiment, this one testing AI Mode rather than AI Overviews, found a forced AI Mode condition cut click-through to external sites and also lowered searchers’ trust, satisfaction and sense of relevance — worst among heavy searchers — while confirming the other randomized study’s finding that removing AI Overviews raises clicks without hurting satisfaction.

The standing answer: It depends on what you publish, and part of the disagreement is instrumentation.

Evidence: study, Search Engine Journal

Still open: Google holds the only complete dataset and has published no methodology behind “billions of clicks”. Until it does, the engine’s own position is the least checkable one in the record.

Read the item →

Now: Google was seen testing the AI Overview on hotel search results moved below the hotel knowledge panel instead of above search results, in screenshots an SEO professional shared and Search Engine Roundtable reported. Google has not confirmed the test.

The standing answer: On the results page itself, and increasingly around it.

Evidence: single source, Search Engine Roundtable

Caveat: single-source — confirmation pending

Still open: Whether the results-page AI Mode button and the local-panel AI Overview treatment ship beyond tests. Google has confirmed neither; each was replicated by one reporter.

Read the item →

Can I buy my way into AI answers?

Before: ChatGPT Ads ran in roughly 40 countries after the August 19 European expansion.

Now: two developments —

  • Two new Adthena readings landed the same day. ChatGPT’s weekly advertiser count rose 65% over eight weeks, from 7,306 to 12,075 — a figure roughly ten times Sensor Tower’s active-advertiser count tracked above, unreconciled. Separately, one- and two-word queries now generate more than half of Google AI Overview ad appearances in every market Adthena measured, reversing a June 2026 pattern where three- and four-word queries led.
  • OpenAI launched ChatGPT Ads in Taiwan, Singapore, Indonesia, Malaysia, the Philippines, Thailand and Vietnam, extending the platform past 60 countries. Ads remain limited to Free and Go plans; businesses buy through self-service Ads Manager or OpenAI’s Ads Solutions team and agency partners including Dentsu, Havas Media, Omnicom Media, Publicis Groupe and WPP.

The standing answer: You can buy a slot, but not a citation — they are separate games.

Evidence: vendor study, PPC Land · company announcement, OpenAI News

Caveat: Adthena study — vendor-published

Still open: Whether buying an ad affects the odds of being cited at all, in either direction. The 3.63% overlap is a snapshot of co-occurrence, not a test — nobody has run the same queries with and without a live campaign.

Read the item → · Read the item →

Now: A Scrunch analysis of nearly 10,000 AI-cited URLs found the first evidence, outside the incumbent-tactics study above, that a site-controlled action moves recommendation: brands whose self-published ‘best of’ or comparison listicle got cited saw recommendation rate rise from about 4% to 7% and outright naming roughly double (19% to 39%). Scrunch’s own falsification test knocked out the study’s riskier secondary finding — that naming a rival can hand it the recommendation — except in a narrow subset, so this adds a data point to the unresolved question below rather than settling it.

The standing answer: Not necessarily — being cited, being named and being recommended are three different outcomes with different drivers, and the record separates them.

Evidence: vendor study, PPC Land

Caveat: Scrunch study — vendor-published

Still open: Whether the 20.8% mismatch holds outside ChatGPT and Semrush’s citation data. It is one analysis of one engine’s estimated demand, and nobody has replicated it elsewhere.

Read the item →

The State of Search keeps one current, sourced answer to each recurring practitioner question. These are the answers this week’s items moved.

Radar

What else moved on the wire this week. None of it changed a standing answer; all of it is on the record.

The patterns behind it

The week’s recurring themes: several items, several subjects, more than one outlet.

Citation measurement

7 items · 9 subjects · 6 outlets · 6 days

Publisher economics

3 items · 6 subjects · 3 outlets · 6 days

AI advertising

3 items · 3 subjects · 3 outlets · 3 days

How we count: a cluster is three or more items on one theme, naming three or more different companies, engines or studies, from at least two outlets, inside the week.

Numbers to keep

From this week’s items, as entered on The Numbers. Evidence type is labelled so a study, a vendor’s measurement, a company’s claim and first-party data read differently.

NumberWhat it measuresEvidence
4% to 7%Lift in a brand’s AI recommendation rate when a self-published ‘best of’ or comparison listicle it wrote gets cited by an AI assistant, versus when it isn’t — across ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, Google AI Overviews and AI Mode.Scrunch study — vendor-published
65%Rise in ChatGPT’s weekly advertiser count over eight weeks, from 7,306 in the week of July 9, 2026 to 12,075 by August 31 — peaking at 13,009 in early August — per an Adthena analysis of AI search queries.Adthena study — vendor-published
>50%Share of Google AI Overview ad appearances generated by one- and two-word search queries, in every market Adthena measured as of late August 2026 …Adthena study — vendor-published

Dates ahead

From the docket: the next dates the record has on file.

  • Oct 30 — In re Mosaic LLM Litigation (Databricks): Hearing before Judge Breyer on whether the case can be decided without a trial (summary judgment)
  • Nov 15 — Bartz v. Anthropic: First class payout, $2,203.56 per work, due on or before this date

The full wire

32 items published September 21–27, every one dated, sourced and linked, grouped by lane, on the week’s page: read the complete wire →

Corrections

None this week. Every correction is logged in public: the corrections log.

From the desk

The week read from the record, not a view of it. The wire is built from a public, dated, sourced record; nothing here is a prediction.

Seen something the record should hold — a platform change, a measurement method, a traffic dataset, a court filing, a publisher policy, a deal term, an ad-product update? Reply to this email. Tips are credited by name or handle, every time.

Archive · State of Search · The Numbers · The Docket · Corrections

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 engineSEOAI visibilityAI OverviewsAI Mode

Via AEO Wire ↗

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

More on the wire

← Back to the wire