Google Downgraded Its Own Search: The Week in SEO, GEO, and AI Visibility September 7 to 11, 2026

The short version

A week of things Google usually does not say out loud

Most weeks in search you are reading tea leaves. Somebody spots a test, somebody else cannot replicate it, a Googler says something carefully neutral.

This week was different. None of the three big stories are leaks or inferences. They are on the record, and together they say something fairly blunt about where measurable search is heading.

Google calls its own EU results the biggest quality drop in 29 years

Google has rolled out the search changes the European Commission required under the Digital Markets Act, and it is not pretending to like them. A Google official told Reuters the changes "mark the largest reduction in quality of service at the world's most popular internet search engine in its 29-year search history."

Mechanically, the changes favor vertical search services. Reuters reported that the revamped results highlight one specialized search engine at the top of the page, followed by two more with fewer details, with a carousel of hotels, airlines or restaurants sitting below them and key features such as real-time prices stripped out. Ranking within that structure is still determined by Google's algorithm.

Practitioners mapping the new layout found the hotel local pack replaced by a multi-provider module featuring the large travel platforms alongside Google, with a separate module surfacing individual hotels and links to their own sites. Hotels have not vanished from the page. They have been moved down a tier and surrounded by intermediaries.

Google has been arguing this since it tested the layout in 2024, when it reported that people were measurably less satisfied, took longer to find hotels, ran more searches, and more often gave up entirely. In that test Google said hotels lost more than 10 percent of traffic while traffic to intermediary sites stayed roughly flat. Google also published a new document covering regional differences in aggregator features.

Why it matters. If you have EU visibility in travel, hospitality or any category with a strong comparison-site layer, your SERP has changed shape and your direct channel has lost ground to intermediaries. Model that before you argue about rankings.

There is also a fair observation worth making, and Search Engine Roundtable's Barry Schwartz made it first: Google is defining quality here by traffic sent to destination sites and by how many searches users have to run. Those are precisely the metrics AI Overviews and AI Mode move in the wrong direction. Google is not wrong that the DMA layout hurts publishers. It is just applying a standard it declines to apply to its own AI surfaces.

People Also Ask is now almost entirely AI Overviews

A SERP feature that has been around for a decade has effectively been replaced from the inside.

Saeed Khosravi of Allintitle reported that Google's People Also Ask answers have been 100 percent AI generated since August 2026. Mark Williams-Cook of AlsoAsked, whose dataset put the figure near 12 percent fourteen months ago, ran a 19.2 million result English language sample and found 97 percent of PAA answers in the first week of September were AI Overviews, up from 86 percent in August. He added that full saturation looks likely shortly.

Both numbers come from commercial SEO tools rather than Google, so treat the exact figure as approximate and the direction as solid. Two independent datasets agreeing on a curve this steep is the useful part.

Why it matters. PAA was the last high-volume SERP feature where a well-structured answer on your page could win a visible, attributable slot. That slot is now an AI Overview, which means the same content competes under grounding and citation rules rather than snippet rules. If you have been optimizing for PAA as a distinct tactic, it has merged into your AI visibility work, and the playbook is the one we laid out in how to optimize content for Google AI Overviews. The practical response is not new content, it is checking whether the pages that used to win PAA are now appearing in the Search Console generative AI performance report.

Google concedes that position 1 to 10 no longer maps

On Thursday, John Mueller answered a question about how impressions are counted for generative AI features in Search Console. He pointed to Google's existing documentation, then added the line that traveled: "the old position 1 to 10 is hard to map."

It is a small comment that carries a lot. Search Console's generative AI performance report went global on August 31 and deliberately reports impressions without position. For twenty years the industry's shared vocabulary has been ordinal: you rank third, you moved up two spots, you are on page one. Inside an AI answer there is no third.

Why it matters. Two things follow. First, if your reporting still leads with average position, it is measuring a shrinking share of reality, and the gap will widen. Second, be skeptical of any tool that hands you a tidy position number for AI surfaces, because Google is telling you it does not have one to give. Presence, citation share and impression trend are the honest units now.

This is harder than it sounds, because position was never just a metric. It was the shared language between agencies and clients, the thing that made progress legible to someone who does not think about search all day. Replacing it means rebuilding how you explain your work, not just which column you export. We wrote about what that reporting looks like in the SEO metrics executives actually need. Teams that start that conversation now will have it on their own terms. Teams that wait will have it during a quarter when the numbers look bad.

Relatedly, Google is still reshaping how those citations look. It spent the week testing a thinner AI Overview citations panel and paginated follow-up responses in AI Mode with a skip button. Neither is confirmed, both point the same direction: less room for links.

Apple separates AI training from Search ranking

Apple quietly updated its Applebot documentation over the weekend, adding a single sentence: "Site rules for Applebot-Extended are not considered in ranking for Search." The Applebot support page now carries that line explicitly.

This mirrors what Google did with Googlebot-Extended. Applebot-Extended governs whether Apple can use your content to train its AI models. Blocking it is a licensing and rights decision. It does not cost you Apple Search visibility.

The timing is not accidental. Days later at Apple's September 9 event, Lilian Rincon, Apple's VP of Product Management and Product Marketing for AI, said Siri now handles 2.5 billion requests every day. AppleInsider reported that the new AI Siri launches in beta with daily usage caps and possible future paid access.

Why it matters. This is the clearest example yet of a distinction most robots.txt files still get wrong. Training access and search access are separate decisions with separate consequences, and the platforms are now saying so in writing. Go and check which of your AI user agent rules are deliberate. Apple has just removed the main excuse for conflating the two, and a surface processing 2.5 billion daily requests is not one to have an accidental policy on. If you want the longer argument for why blocking without scarcity is not a strategy, it is in Reddit's paywall doesn't control distribution.

Google Ads is clawing back credits after advertisers spend them

The week's ugliest story is a billing one. PPC consultant David Melamed reported that Google Ads invalidated a 3,200 dollar promotional credit more than a month after his client had already spent against it, and that it was the second account this had happened to.

His complaint is specific: the credit is marketed to induce spend, there is no known appeal route once it is invalidated, and the spend itself cannot be recovered. He argued the client would not have spent the first 3,200 dollars without the credit. Google Ads Liaison Ginny Marvin replied publicly that she had passed it to the team.

Why it matters. If you run accounts with promotional credits, reconcile them. Do not treat an applied credit as banked until it has survived past the point where it could be invalidated, and flag the risk to clients before you let a credit shape a spending decision. Melamed also makes a point worth sitting with: credit-driven spend inflates auction prices for everyone bidding alongside it, whether or not the credit ultimately holds.

Also worth knowing

Scammers found a new route into local. Search Engine Roundtable reported a Google Maps photo scam in which the contributor program is used to upload AI-generated images containing a phone number, so callers dial the number in the photo rather than the verified one on the listing. It is running across dozens and possibly hundreds of profiles, and reporting it appeared to achieve little. Check your own listing photos.

A new ranking factors survey landed. Cyrus Shepard and Dawn Shepard published the 2026 expert survey at Zyppy, covered by Search Engine Land with 131 SEO professionals responding and by Search Engine Roundtable. Content relevance, backlinks and content quality came out on top. It is practitioner opinion rather than measurement, which is worth remembering, but the consensus is notably unexciting and that is probably the finding.

Google Merchant Center gained AI performance reporting covering AI search intent, terms and attribution, which is more query-level AI data than Search Console currently offers. If you sell products, that report is worth opening this week, because it is currently the most granular view Google gives anyone into which AI queries touch their listings.

Google also warned that it can lose faith in sites publishing low value programmatic content at scale, a useful counterweight to the idea that AI makes volume cheap enough to be worth it. The Search Console links report has not updated in about a month, so do not read anything into flat link counts. Google also confirmed an indexing reporting blip and Business Profile insights were missing September data, both covered in Barry Schwartz's daily recaps.

What to actually do with this week