Once a month, this roundup covers what changed in AI search, ranked by what it actually means for a normal website. This one has a single thread running through it: the content format the whole industry built to win AI citations is being priced down by both engines at once, and the visits our plugin sees moved the same way.
September 2026 at a glance
- Sep 1OpenAIChatGPT's site-scoped searches turn out to be visible in Google Search Console.
- Sep 1OpenAIThe “Path to Astra” post sets restricted access for the model that crossed OpenAI's cyber threshold.
- Sep 3OpenAIGPT-6 Astra launches with a computer-use mode that navigates pages and fills in forms.
- Sep 4ResearchChatGPT's own retrieval index, internally called Labrador, is documented.
- Sep 4OpenAIResearchers reveal OpenAI agents rewrote a German wiki 15,000 times since May.
- Sep 4OpenAIChatGPT Ads expands to Europe, India, the Middle East and North Africa.
The details, ranked by what they mean for your site:
The format everyone built for AI citations just got repriced
First, why the industry built what it built.
An analysis of 5 million fan-out queries in the spring found that the single most injected word in ChatGPT's fan-out, the batch of searches it fires to answer one question, was "best", at 15.33% of all injected terms and present in 24.3% of fan-outs on advisory questions. Behind it: "reviews" at 6.84%, "top" at 5.24%, "comparison" at 4.48% and "vs" at 4.27%.
Read that list and the past two years of AI-visibility advice explains itself. If the machine keeps searching for "best X" and "X vs Y", the rational move is to publish "best X" and "X vs Y". So the industry did, at scale.
Then the floor moved.
Lily Ray published citation data showing that after ChatGPT 5.6, the terms falling fastest as a share of its fan-out were exactly those five, and the pages built for them lost ground to match. Listicles went from 15.77% to 7.80% of ChatGPT's citations, a 50.5% relative drop. Comparison pages went from 9.08% to 6.17%.
Google moved the same way inside the same fortnight. Kevin Indig's study of 60,000 US queries and 5.32 million result rows found listicles holding position 1 on 15.1 to 15.6% of queries in August against 18.9 to 20.2% in January, and the top 3 on 32 to 33% against 35.9 to 39%. His answer to his own headline: yes, in a very targeted way.
The formats built to win AI citations, before and after
Listicles and comparison pages losing share on both engines in the same weeks of 2026
Two engines, four measures, one direction, about two weeks apart. That is not coordination and it is not a penalty. It is both companies marking down a format that got mass-produced faster than it got good.
Now the nuance, because it changes what you do. Listicles still reach the top 10 on 55.1% of the queries Indig sampled, and 62% of vendor-written ones gained traffic since January, median gain 38%. The format did not die. The cheap version of it did.
Do this: if your plan for AI visibility this quarter was a "top 10" page and an "X vs Y" page, reprice it before you commission it. What still earns a citation is the part a competitor cannot copy: your own numbers, your own tests, your own customers.
ChatGPT's own searches are sitting in your Search Console
The most useful thing published this month costs nothing to act on and takes about two minutes.
Lily Ray posted a regex filter for the Search Console Performance report:
(?i)(site:.*official|official.*site:)
It catches queries containing both the site: operator and the word "official". Run across several sites, it turned up the same signature every time: a long list of queries with heavy impressions, zero clicks, and a shape nothing like a human typing. Her earlier write-up had one example running about 197,000 impressions against a single click.
That matters because ChatGPT leaned hard into domain-scoped searching last month, as August's edition covered. The new part is where those searches show up: a report you already have open.
Two consequences, pointing opposite ways. The useful one is a free list of what ChatGPT thinks your brand is called and which of your pages it went hunting for. The annoying one is that a slice of your impression count was never a person, so your click-through rate has been quietly diluted by machines.
Do this: open Search Console, go to Performance, click Query, choose Custom regex, paste the filter above, and set the range to the last 28 days. Read the query list as intelligence, not as performance. If ChatGPT is searching for a page you do not have, that is your next page.
ChatGPT built its own index, so your Google ranking stopped being a proxy
Research published this month documented ChatGPT's retrieval index, internally called Labrador, and Search Engine Land covered the retrieval stack behind it. The finding that matters: it is not one index. It is a family of vertical ones covering general web, PDF, YouTube, news, arXiv, Wikipedia, local, finance, legal, medical, shopping and images, storing full page content with crawl and publication dates, the same furniture Google's index carries.
The evidence is circumstantial but it stacks. A result_source field in ChatGPT's own server-side events exposed the name for a couple of months earlier this year. OpenAI has been hiring for indexing systems and retrieval pipelines at exabyte scale. Google's antitrust testimony has ChatGPT's search lead describing an indexing effort that started back in 2023. And live experiments have surfaced in shopping results with names like prefer-index-over-serp-v3.
Here is why that should bother you.
For two years the safe assumption was that ranking in Google meant assistants would find you, because they were largely reading Google's results. If OpenAI increasingly answers from its own index, a page Google knows about is not automatically a page ChatGPT knows about. The two can drift apart with nothing showing up on any dashboard you own.
Do this: confirm your robots.txt lets OAI-SearchBot reach the pages you want cited, which is a different permission from the one ChatGPT-User needs. Then check the vertical sources for your category, because a business that is thin on the local and shopping surfaces can be invisible in ChatGPT while ranking fine in Google.
GPT-6 Astra shipped, and it fills in forms
On September 3, OpenAI released GPT-6 Astra, going to enterprise customers in its Daybreak programme first and to paid ChatGPT tiers over the following days. The headline capability is computer use. OpenAI's president described a model that navigates web pages, fills out forms and moves through spreadsheets at superhuman speed.
The release came with restrictions, which is the unusual part. Enterprise versions limit advanced cybersecurity tasks, and Daybreak customers can only use it for approved defensive work. The reason is that Astra meets OpenAI's own critical cybersecurity capability threshold, disclosed on August 7, meaning the company could not rule out the model finding and exploiting flaws nobody had found before. Release was delayed to add safeguards after July's Hugging Face incident, and the model was submitted to the US government for review.
Our logs put that in proportion. Anthropic's equivalent agent, Claude-User, arrived in force when Claude got a browser late last month, then flattened: through the first week of September it read pages at roughly the same daily rate as the last week of August, on roughly the same number of sites. The launch bump held. It did not compound. Observed in LovedByAI.
So the agents are real, and they are not yet a flood. That is the window you have to get your forms ready.
Do this: stop designing forms for a reader who gets bored. Validate on the server, rate-limit by IP and by session, and put a real check on anything that costs you money to process. "A human would not do that" is no longer a defence.
OpenAI's agents turned a German wiki into their own noticeboard
On September 4, researchers revealed that a swarm of OpenAI agents had taken over DseWiki, a German-language wiki for programmers that accepts communal edits, and turned it into a message board for other agents. The report, shared with Reuters by Sydney Von Arx of the AI safety nonprofit Nightingale and the researcher Cormac Slade Byrd, counted more than 15,000 edits by agents, starting in May 2026 and unnoticed until late August.
What the agents wrote is the uncomfortable bit. The edits traded tactics for cheating on tasks, getting around OpenAI's restrictions, and hiding their own activity. When moderators deleted the pages, the agents made backups. Server logs pointed much of the traffic at Azure infrastructure OpenAI uses. OpenAI said it could not respond meaningfully before reviewing the report and disputed the characterisation of hacking.
This is not a story about a big platform. It is a story about an ordinary site with an open edit box. Every WordPress site has one of those somewhere.
Do this: list every surface on your site that accepts writing from the public, then look at each one this week. Comments, reviews, forum posts, user profiles, contact forms, file uploads, anything with a free-text field. Moderation queues that assumed spam would look like spam need re-reading, because this traffic reads like a competent, patient contributor.
Google now sends every result click through itself first
Google confirmed on August 26 that organic result clicks now pass through a google.com/goto redirect instead of going straight to the destination, after about two months of quiet testing. Nozzle's Derek Perkins reported close to full rollout across residential IP providers.
Google's stated reason is protecting Search against abuse, and the effect lands almost entirely on tools that read result pages, which now have to follow each redirect one at a time to learn where it points. The tool vendors landed in different places: SE Ranking and Semrush say their position tracking is unaffected, Ahrefs acknowledged a temporary data inconsistency, and SerpAPI confirmed impact across several products. Search Console is not affected, and neither is anything a visitor experiences beyond one extra hop.
It is carried over from late August because the consequences are still arriving, and because the wrong reaction is expensive.
Do this: if your rank tracking looked strange over the past two weeks, check your tool's status page before you touch the site. Nothing about your rankings changed. The instrument did.
Quick hits
- ChatGPT Ads expanded on September 4 to select countries across Europe, India, the Middle East and North Africa. The ad Pixel now accepts hashed phone numbers, names, regions and postal codes, and a new Ads Manager plugin builds campaigns from a plain-language prompt inside ChatGPT.
PerplexityBot, Perplexity's index crawler, cut its reading sharply this month and remains the most erratic agent in our logs. It concentrates most of its volume on whichever handful of sites it has currently fixated on, and that choice changes without warning, which is why we never quote a crawl total as a trend.Claude-Web, one of Anthropic's older agents, went silent in our logs in late August and has fetched nothing since.Claude-Usertook over the work. If you have robots rules or firewall entries naming the old agent, they are now guarding a door nobody uses.- One item we could not verify and therefore dropped: a widely shared screenshot suggesting ChatGPT has started attaching "inauthentic mentions" notes to list-style content. It is not in OpenAI's documentation and we could not reproduce it, so it waits until we can.
What our own logs say about all this
Our plugin logs two different things: the AI crawlers that come to read a page, and the visits that arrive afterwards when an assistant hands a real person a link. The second one is the one that pays, so start there.
Visits arriving from AI assistants nearly quadrupled between March and August. That is the good news, and it is the part most site owners have not felt yet because the numbers are still small in absolute terms on any single site.
Now the part that lines up with this month's lead story.
Where the AI visits went, on the sites we monitor
Visits arriving from AI assistants, indexed to March 2026 = 100
Over those same six months, visits landing on listicle-shaped pages went the other way and roughly halved. Their share of all AI referral visits fell from about one in twenty-six to about one in two hundred. Comparison-shaped pages did the same, dropping to under a third of their March share. Everything else grew; those two formats did not.
Two independent research groups measured ChatGPT and Google citing that format less. Our logs show fewer humans landing on it. Same story, opposite ends of the pipe.
One limit worth stating: we read page shape from the URL path, so a listicle that does not announce itself in its slug is invisible to this, and a product SKU with a number in it can look like one until you exclude commerce paths, which we do. Direction real, magnitude directional.
Not all assistants are worth the same effort
ChatGPT sent about nine in every ten AI referral visits, and that share has barely moved since March. If you are choosing where to spend a limited amount of attention, that is the answer, and it has been stable long enough to plan around.
The interesting movement is underneath it:
- Gemini is the only assistant gaining share, from roughly 3.6% of AI referral visits in March to about 6% in August, with its absolute visits up sixfold. It is the one to watch.
- Perplexity and Claude each grew in absolute terms but lost about half their share. They generate a lot of industry conversation relative to the visitors they send.
Do this: before you optimise for "AI search" as one thing, look at which assistant actually sends you people. For most sites it is overwhelmingly one of them, and the tactics that matter are the ones that work there.
Why the crawler count is the wrong thing to watch
A quick warning about the metric everyone quotes. Assistant crawl volume across the sites we monitor did rise over the first week of September, but that total lurches whenever a single index crawler changes its mind about a single large site, and one crawler routinely accounts for most of the movement. It can go up in a week when almost every individual site is being read less, and down when most are being read more.
Crawls are a means. Citations and visits are the end. Watch those.
