AI Search News: September 2026

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. For the longer-run picture behind these monthly readings, our 2026 AI traffic study benchmarks what AI assistants actually send to small business sites over a full window.

September 2026 at a glance

1. Sep 1OpenAIChatGPT's site-scoped searches turn out to be visible in Google Search Console.
2. Sep 1OpenAIThe “Path to Astra” post sets restricted access for the model that crossed OpenAI's cyber threshold.
3. Sep 3OpenAIGPT-6 Astra launches with a computer-use mode that navigates pages and fills in forms.
4. Sep 4ResearchChatGPT's own retrieval index, internally called Labrador, is documented.
5. Sep 4OpenAIResearchers reveal OpenAI agents rewrote a German wiki 15,000 times since May.
6. Sep 4OpenAIChatGPT Ads expands to Europe, India, the Middle East and North Africa.
7. Sep 8OpenAIGoogle Shopping URLs start showing up as named sources inside ChatGPT.
8. Sep 8OpenAIChatGPT Images 2.5 lands, with sharper detail and more precise editing.
9. Sep 9ResearchAlsoAsked finds 97% of People Also Ask answers are now AI Overviews, up from 12% fourteen months earlier.
10. Sep 9GoogleMerchant Center AI insights adds AI search intent, AI search terms and AI attributes.
11. Sep 10OpenAIA Data plugin arrives for ChatGPT Work and Codex; Box, Dropbox and SharePoint join Library.
12. Sep 12GooglePractitioners report a large unannounced movement in Google's rankings. Google has not confirmed it.
13. Sep 14GoogleGoogle starts testing AI contribution, a Search Console panel that pays sites whose content feeds an AI answer.
14. Sep 15ResearchA randomised experiment finds AI Overviews cut clicks by nearly 40% without improving satisfaction, and that only 7.1% of the remaining clicks come from citations inside the answer.
15. Sep 17–18GoogleFree product listings and product carousels disappear across the European Economic Area under the Digital Markets Act.
16. Sep 18ResearchChatGPT's retrieval is shown to buy results from scraping vendors alongside its own Labrador index, and to license Yelp and TripAdvisor.
17. Sep 19GoogleGoogle's local business knowledge panel is spotted rendering as an AI Overview.
18. Sep 22GoogleGoogle auto-enrols eligible US Shopify stores in native checkout inside AI Mode and Gemini.
19. Sep 24GoogleThe September 2026 spam update starts rolling out, with up to two weeks to complete.
20. Sep 24GoogleSearch Console splits performance into text-based and multimodal searches.
21. Sep 28ShopifyShopify opens checkout to browser-based AI agents, which can now submit an order once the buyer approves.
22. Sep 29ResearchA 460,000-answer study finds AI shopping assistants pick a national chain over a smaller store up to 94% of the time.
23. Sep 30GoogleAI contribution payments are reported at under 0.1% of ad revenue for smaller sites in the pilot.
24. Sep 30GoogleA US judge dismisses the Chegg and Penske Media lawsuits over AI Overviews.

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.

The month closed with a third reading pointing the same way. On September 30 Lily Ray re-ran a check of about 100 queries and found Google's AI Overviews citing 38% fewer self-promotional "best of" listicles than in June. When they did cite one, they left the brand behind it out of the answer 83% of the time, up from 69%. In 70% of cases the brand's own listicle was helping its competitors get recommended while the brand itself was not.

That last number is the one to sit with. A "best X" page that puts your own product first is now, more often than not, a page that sends the AI to your rivals. The sample is small, so treat the exact figures as directional, but the direction matches everything else in this section.

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.

Being cited inside an AI Overview returns almost none of the clicks

The story above says the formats got repriced. This one says the prize itself is smaller than the industry has been selling, and it is the most important thing published all month.

Saharsh Agarwal and Ananya Sen ran a randomised experiment on 1,065 US desktop users, written up by Digital Content Next on September 15. Everyone searched Google normally; one group had AI Overviews stripped out by a browser extension, with the ordinary results left alone. That design is what makes it worth reading, because it compares the same people doing the same searches with and without the box.

Three findings, and the third is the one to write down.

Removing AI Overviews took outbound organic clicks from 0.37 to 0.62 per search, so the box costs publishers close to 40% of their clicks. Zero-click searches ran at 73% with AI Overviews against 54% without.

Search satisfaction did not move. It sat at 4.0 out of 5 either way, and there was no significant difference in time spent on the destination site, bounce rate, or how often people came back to Google. So the clicks AI Overviews suppress were not low-quality clicks being filtered out. They were simply clicks that stopped happening.

Then the number that should change what you spend money on. When someone does click through from a page carrying an AI Overview, only 7.1% of those clicks come from the citation links inside the AI answer. The other 92.9% come from the ordinary blue links underneath it.

Read that against two years of advice. The whole GEO industry has been selling a place inside the answer box. On this evidence a place inside the box is worth about one click in fourteen, and the traditional ranking below it is still doing nearly all the work.

Do this: keep your classic ranking work funded. If a vendor is pitching you on citations alone, ask them what it is worth in visits, and hold the 7.1% figure next to their answer. Being cited is a brand impression. The link below the box is still where the visitor comes from.

Google's September 2026 spam update is rolling out

The only confirmed Google ranking update of the month landed late. Google released the September 2026 spam update on September 24, applying globally and in all languages, with a rollout that may take up to two weeks. That puts its end around October 8, so some of the effect is still arriving as you read this.

Google called it a normal spam update and named no new technique as its target. It is the fourth spam update of 2026, and a slower one: Barry Schwartz noted that August's spam update finished in two days, and John Mueller said this one would take longer than some previous ones. Trackers reported ranking volatility on September 23 and 24, around the start of the rollout.

Here is why the timing matters. The movement practitioners saw around September 11 and 12, covered in the Quick hits, was never confirmed. This one is. If your traffic changed in the last week of September, this is the first thing to rule in or out.

Do this: mark September 24 in your analytics and compare the two weeks either side once the rollout finishes. If you lost ground, read the page that drew it against Google's spam policies before you touch anything else: scaled thin pages, expired domains bought for their links, and third-party content parked on your site are the usual suspects. Google's own guidance is that recovery can take months, so a fast fix is not the goal. A correct one is.

Google is testing paying publishers when its AI uses their content

The other side of the AI Overview click problem arrived a day before that experiment, and you may already be enrolled without noticing.

On September 14 it emerged that Google is piloting a Search Console panel called AI contribution, which pays sites whose content is used while Gemini, AI Overviews and AI Mode generate an answer. Google confirmed the pilot is early-stage and pointed back to a June 18 post about rewarding content that meaningfully contributes to the freshness and factuality of generative answers.

The mechanics are thin on purpose. The panel shows a monthly earnings figure and some history. It does not show which pages earned the money, how often your content contributed, or how the number was calculated. One publisher executive quoted in the reporting called it "quite black box", and another described the early sums as "peanuts" next to ad revenue.

Two things are worth separating. The payments are small and opaque, so this is not a revenue plan. But the existence of the panel is the interesting part: it means Google is now measuring, per site, whether your content shaped an AI answer rather than merely got linked afterwards. That measurement only counts contributions at the generation stage. Being cited underneath, or used to check a fact, does not qualify.

The money got a number at month-end. Reporting by The Information, published on September 30, put the pilot at roughly 100 publishers, with several small and midsize sites earning less than a tenth of one percent of their advertising revenue from it, and the smallest under $1,000 over several months. One early participant was said to be earning more than $1 million a year. So the payments scale with size, and for a small site they are not yet worth planning around.

Do this: open Search Console and look for an AI contribution panel. If it is there, you are in the pilot, and the monthly figure is the only signal anyone currently has about whether Google thinks your content feeds its answers. Do not restructure anything around it yet. Reporting this vague is not a strategy input.

People Also Ask has finished turning into AI Overviews

The other format that fed ordinary sites clicks closed this month, and almost nobody announced it.

Mark Williams-Cook ran the question against AlsoAsked's dataset of 19.2 million English queries sampled through 2026 and reported on September 9 that 97% of People Also Ask answers in the first week of September were AI Overviews. In August it was 86%. Fourteen months ago it was about 12%. Saeed Khosravi of Allintitle ran a separate sample with the Also Ask Miner tool and found 100% of the August answers he checked were AI generated.

Williams-Cook's own read: it seems highly likely the number reaches 100% shortly.

Worth being precise about what closed. People Also Ask was never a huge traffic source, but it was a predictable one, and it was the surface where a small site could out-answer a big one on a narrow question. That box now resolves into an AI answer, so the question still gets asked and the click mostly does not arrive.

The useful half is that the questions themselves are unchanged. PAA remains the cheapest public map of how real people phrase a problem in your category. It has simply stopped being a place you win a visit and become a place you win a mention.

Do this: keep mining PAA questions for what to write about, and stop counting on them for clicks. If a page of yours exists mainly to capture a PAA box, check its traffic over the last fourteen months before you spend anything else on it.

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.

A follow-up published on September 18 by the researcher Tomek Rudzki sharpened the picture, and it complicates the word "own". Reading the same server-side events between May 21 and July 21, he found the result_source field only ever carried four values: Labrador, Bright, Oxylabs and SERP. Two of those are commercial scraping vendors. So ChatGPT is not simply reading its own index, it is also buying search results wholesale. Alongside them sat licensed vertical sources, including Yelp, TripAdvisor and Microsoft's Web IQ.

That is good news and bad news in one finding. The bad news is that there is no single door to get through. The good news is that some of those doors are ones you can walk up to, because a Yelp or TripAdvisor listing is something you control directly and most businesses last updated years ago.

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. If you are a local or hospitality business, treat your Yelp and TripAdvisor entries as retrieval surface, not as legacy directories.

Google will now tell you which AI search terms your products are missing

If you sell anything, this is the most concrete thing that happened all month.

Google's AI Performance Insights report in Merchant Center gained three new sections, reported on September 9 and expanding to more regions through the month after a July launch. In Google's own wording, the report now covers:

• AI search intent, which shows how your products match customer AI searches.
• AI search terms, which recommends popular terms to include in your product descriptions for products already showing up in AI searches.
• AI attributes, which recommends missing attributes to add to those same products.

Read the second and third of those again. Google is naming the words it expects on your product pages and the structured fields it thinks you left blank, for the products it is already surfacing in AI answers. That is the kind of thing the AI visibility industry charges for, and it is sitting inside a free tool most merchants already have connected.

There is a reason to act on it beyond Google. Practitioners spent the second week of September reporting Google Shopping URLs turning up as named sources inside ChatGPT, which is covered in the Quick hits below. If that holds, your product feed is not only a Google asset. It is one of the routes into both engines.

Do this: open Merchant Center, find AI Performance Insights, and work the AI search terms and AI attributes lists. Fix the products it says are already appearing first, because those are the ones where a missing attribute is costing you a sale you were otherwise going to make. If the report is not in your account yet, it is still rolling out by region.

Checkout moved inside the AI assistant, and the assistants prefer big chains

Three things happened in the last ten days of September that, read together, change where an online sale can happen.

On September 22, Google started emailing US Shopify merchants that their products were now eligible for native checkout, meaning customers can buy them inside AI Mode on Google Search and in the Gemini app without visiting the store. Eligible products were switched on automatically. In Google's words, there was nothing to set up.

On September 28, Shopify opened its checkout to browser-based AI agents through WebMCP. An agent can now read the checkout, change the delivery address or shipping option, and submit the order once the buyer has authorised it, rolling out to all eligible merchants. Amazon and Adidas are going the other way and blocking agents from buying.

Then on September 29, a Lightspeed Commerce study of 460,000 AI responses across ChatGPT, AI Mode, AI Overviews and Gemini found the assistants chose a national retailer over a smaller competitor up to 94% of the time in head-to-head shopping recommendations. Even with live web search on, national chains took 46 to 58% of recommendations, and 68% of AI Overviews in the test carried no local store at all.

Here is the tension. The assistant can now complete the sale for you, and the same assistant is, by default, recommending someone bigger. Being buyable inside the answer is worth nothing if you are not in the answer.

Do this: if you are on Shopify, open Sales channels, then Agentic, and decide on purpose whether native checkout stays on, because it is already on. Then make the case for your store in words an assistant can lift: where you are, what you stock that a chain does not, and why someone local would choose you. The study found that one word in the shopper's prompt, "independent", cut chain recommendations from about 44% to as little as 9%, so give the assistant the facts that let it answer that question about you.

Free product listings vanished across Europe in two days

And then, for European merchants, the free surface the Merchant Center advice applies to largely disappeared.

Over September 17 and 18, Google removed free product listings and the popular-products carousels from search results across the European Economic Area, with no announcement and no transition period. Hugo Huijer of Productrise measured drops of roughly 90 to 100% in Germany, France, Belgium, Sweden and the Netherlands, and Semrush tracking showed the same. Ginny Marvin, Google's Ads Liaison, confirmed the removal follows from Digital Markets Act requirements. The compliance deadline in the relevant case falls on September 21.

This is the end of a chain that has been running all year. The European Commission fined Google €460 million in July for giving its own shopping and travel results unfair prominence, and Google rebuilt EEA results on September 8 around aggregator and supplier units that give comparison services their own placement. Nick Fox, Google's SVP of Knowledge and Information, called the wider set of changes the largest reduction in quality of service in Search's 29-year history, and said Google expects them to hurt local businesses in favour of intermediaries. Google has previously reported that earlier DMA compliance work cut free direct booking traffic to European businesses by 30%.

Set the self-interest in that statement aside and the practical position is unchanged: if you sell to Europe, the free organic route to product visibility in Google is now mostly gone. Paid Shopping ads are untouched. The space went to comparison shopping services. So the two remaining routes are a CSS or a credit card.

Do this: if you sell into the EEA, check your Search Console and analytics for a step change on September 17 and 18 before you go looking for a ranking problem, because this one is structural and nothing you did caused it. Then price the two real options: joining a comparison shopping service, or paying for Shopping ads. If most of your European product traffic was free listings, that budget line is now compulsory rather than optional.

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.

More AI search news from September 2026

• ChatGPT Ads expanded on September 4 to select countries across Europe, India, the Middle East and North Africa, and on September 23 added seven Asian markets to pass 60 countries. The ad Pixel now accepts hashed phone numbers, names, regions and postal codes, a new Ads Manager plugin builds campaigns from a plain-language prompt, and on September 30 advertisers got bulk creation for product feed campaigns.
• Google is testing AI Overview links that open AI Mode instead of a website. Gagan Ghotra spotted anchor links inside AI Overviews on September 22 that take the searcher into an AI Mode follow-up rather than to a page. A day later, tracking data shared by Malte Landwehr showed the share of AI Overviews carrying an external link in the answer text jumping from almost none to over a quarter. Both are moving, in opposite directions, so watch your AI Overview clicks rather than assuming either.
• Search Console now separates multimodal searches. Since September 24 the Performance report has a Web multimodal search type for results triggered by an image, photo or screenshot, covering Google Lens and Circle to Search. There is no query data for it, but if you sell anything visual, it is the first place you can see that traffic on its own.
• PerplexityBot, Perplexity's index crawler, is the month's best example of why one week is not a trend. It cut its reading sharply in the first week of September, which we reported here, then grew every week after, and over the full month the typical site was read 55% more than in August. It also spread out: in August one site took 78% of its reading, in September the top site took 19%. Same crawler, opposite story, three weeks apart.
• Claude-Web, one of Anthropic's older agents, went quiet in our logs in late August and Claude-User took over the work. It is not quite retired: it reappeared from September 17 and kept trickling through the month, one site per day across a handful of sites, for 48 requests in all. That is small enough to be a leftover rather than a comeback, but if you wrote robots rules or firewall entries naming the old agent, they are still doing something.
• Google's local business knowledge panel is being tested as an AI Overview. Barry Schwartz reported the box that normally carries a company's hours, phone number and reviews appearing instead as an AI Overview that describes the business in Google's own words, with a Show more button leading into an AI Mode chat. It is a test, not a rollout, but it is the panel most local businesses rely on.
• Google Discover is testing a "Dive deeper" button that opens an AI-generated overview rather than the publisher's article, reported September 18. Lily Ray's reaction was that Google seems never satisfied in taking traffic from publishers. Chartbeat's publisher data, reported September 24, already had Discover referrals down 34.3% year on year. If Discover is a real referrer for you, watch it.
• Google Shopping URLs are showing up as sources inside ChatGPT. David Konitzny reported seeing Google named as a source in ChatGPT running GPT-6 Astra on September 8, and Lily Ray reported more Google Shopping URLs in ChatGPT's network response on September 12. Two named practitioners reading the underlying responses rather than passing round a screenshot, but still observation rather than study, so treat it as a lead worth watching rather than a settled fact. It is the reason the Merchant Center item above matters outside Google.
• Rankings moved around September 11 and 12, and Google never said why. Glenn Gabe flagged the movement, Marie Haynes confirmed seeing it on September 12, and Lily Ray was asking on September 13 whether something bigger was coming. The month ended with no confirmation and no tracking vendor dataset, and with the spam update above as Google's only confirmed ranking change of September. Treat the mid-month movement as unexplained, and do not attribute it to the spam update, which started twelve days later.
• Google Maps now asks people to sign in to read all of a business's reviews, reported September 30, along with sorting and posting them. Barry Schwartz's reading was that it is aimed at bots and scrapers. For a local business it means a logged-out visitor, and any AI tool reading the page, now sees a shorter version of your reputation.
• A US judge dismissed the Chegg and Penske Media lawsuits over AI Overviews on September 30. Judge Amit Mehta treated AI Overviews as an improvement to search rather than a separate product tied to it. Both rulings can be appealed, but for now the courts are not going to give publishers their clicks back.
• Google Search Console's crawl stats report dropped a day of data again, this time September 15. It has happened repeatedly and Google has always restored the data, so treat the gap as a reporting fault rather than a change in how often Google visits you.
• OpenAI shipped ChatGPT Images 2.5 with sharper detail and more precise editing on September 8, and on September 10 added a Data plugin for ChatGPT Work and Codex plus Box, Dropbox and SharePoint alongside Google Drive in Library. At month-end it pulled the next Astra release, GPT-6.1 Astra, after testing found it overstepped its instructions and misled users about what it was doing. None of these changes anything a site owner does.
• 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.

AI referral and crawler data for September 2026: what our logs say

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.

On the sites we monitored in both months, visits arriving from AI assistants more than doubled between March and September. 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 the six months to August, 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. September did not reverse it: AI referral visits rose by about a quarter on August while visits to listicle-shaped pages slipped again.

Two independent research groups measured ChatGPT and Google citing that format less, and Lily Ray's month-end check found AI Overviews doing it too. 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.

ChatGPT read less of every site and sent more people anyway

This is the number that moved most between the third week and the end of the month, and it is the most useful thing our logs said all September.

ChatGPT-User is the agent that fetches a page live because someone asked ChatGPT something. On the same fixed set of sites, it fetched the typical site 41% less in September than in August. That alone is not news: August included an eleven-day surge, covered in the August edition. The news is that the pullback never bottomed. Measured against the last eleven days of August, after the surge had already faded, each stretch of September came in lower than the one before, and the last ten days ran 41% below late August.

The number of sites it touched barely moved: the median day reached about 4% fewer sites at the end of September than at the end of August. So it is not abandoning sites. It is fetching far fewer pages from each one.

Now put that next to the visits. On the same sites, ChatGPT's referral visits rose 29% in September.

Fewer live reads, more people sent. That is what you would expect if ChatGPT is answering more from its own stored copy of the web, the Labrador index described above, and checking your server live less often. It has a practical edge: if your only signal for ChatGPT's interest is the crawler log, you would have concluded it lost interest in you this month, and you would have been wrong.

Do this: judge ChatGPT by the visits it sends, not by how often it fetches. And because more of its answers now come from a stored copy, make sure the copy it has is a good one: keep OAI-SearchBot allowed, and keep the facts on your key pages current, because a stale page can now sit in an answer for longer before a live fetch corrects it.

Not all assistants are worth the same effort

Among the assistants that identify themselves, ChatGPT sends the most visitors by a wide margin, and in September it pulled further ahead: on the same sites its referral visits rose 29%. Treat any "share of AI traffic" figure with care, ours included. A click from a Google AI Overview or from AI Mode arrives as a plain google.com visit, so it sits in your organic line and no log can split it out. ChatGPT also tags its links, which makes it easier to count than anyone else.

The movement underneath it:

• Gemini grew fastest through the spring and summer, with its referral visits up sixfold between March and August. In September its climb stalled: on the same sites its visits slipped about 4% while ChatGPT's rose 29%. One month is not a reversal, and the drop is spread across many sites rather than one, so it is still the one to watch.
• Perplexity and Claude both grew over the six months, but from much smaller bases and more slowly than ChatGPT. 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, and remember that Google's AI answers are hiding inside your organic search numbers. The tactics that matter are the ones that work where your visitors really come from.

One week is not a trend, twice over

Two agents made the same point this month from opposite directions.

Anthropic's Claude-User widened sharply in the second week of September, reaching about 40% more sites a day than in the first week, and we reported that here. In the third week it gave most of the widening back, and in the last ten days it settled about a fifth above where it sat at the end of August. A bump, then a modest step, not a surge.

PerplexityBot went the other way. Its first week of September looked like a sharp cut, and we reported that too. Over the full month the typical site was read 55% more than in August, and its reading spread from one dominant site to many. Had we stopped at the first week, we would have told you the opposite of what happened.

Why the crawler count is the wrong thing to watch

A quick warning about the metric everyone quotes, and this month our own data makes the case better than an argument could.

Take one fixed set of sites and one honest question: did assistant crawling rise or fall in September? Measured over the first six days against the same six days of August, the total was up 9.4%. Extend both windows to thirteen days and it is down 5.9%. Twenty days: down 8.0%. The full month: down 27.8%. Same sites, same bots, same method, four answers, and the first one had the wrong sign.

Here is why. One index crawler accounts for roughly two in three of all the assistant fetches we see, and about three quarters of its September reading landed on a single site. So the fleet total is, in practice, that one crawler's current opinion of that one site. When it loses interest the industry metric falls, and almost nothing has happened to anyone else.

The median tells the sober version: across sites with a real August baseline, the typical site was read 13.1% less in September, with 57% of those sites down and 43% up. A mild, broad easing, driven mostly by ChatGPT's live fetcher, and accompanied by more visits rather than fewer.

Crawls are a means. Citations and visits are the end. Watch those. If you want to know where your own site sits on that measure, the free GEO checker reads a live URL, our on-page GEO guide covers what to fix in order, and last month's edition covers what changed before this. October's edition picks up from there.