What Actually Changed in SEO in 2026

1. Ranking and citation came apart

For most of 2025 the working assumption was straightforward: get into the top 10, and you were in the pool that AI Overviews drew from. That assumption no longer holds.

Ahrefs analysed 863,000 keyword SERPs and roughly 4 million AI Overview URLs in January 2026. Only 38% of AI Overview citations came from pages ranking in the top 10 — down from around 76% six months earlier. Looking at organic results alone, 37.1% came from the top 10, 26.2% from positions 11–100, and 36.7% from pages that did not rank in the top 100 at all.

The mechanism is something Google now documents openly in Search Central: query fan-out. Rather than answering the query you typed, AI Overviews and AI Mode issue a spread of related sub-queries and assemble the answer from all of them — which Google says lets it "display a wider and more diverse set of helpful links" than classic results.

This reframes the work. Ranking position is now one input rather than the gate, and what determines whether you enter the candidate pool for an unseen sub-query is coverage: whether your site addresses that specific sub-question at all, in text a machine can extract.

2. The click numbers everyone quotes — and the one nobody does

The headline figure of 2026 is SparkToro's: analysing US clickstream data from January to April 2026, 68.01% of Google searches ended without a click, up from 60.45% in 2024 and roughly 45% in 2016. Their methodology excludes Google's mobile search app, so the true figure is likely higher.

On click-through rate, the range is wide enough that any single number deserves suspicion. Ahrefs measured a 58% lower average CTR for top-ranking pages when an AI Overview is present — comparing December 2023 with December 2025, a two-year baseline gap they disclose as a compromise. Authoritas measured 47.5% on desktop. Ahrefs themselves note published estimates range from 47.5% to 90%.

Then there is the datapoint almost nobody reports. Seer Interactive tracked 53 brands and 5.47 million queries from January 2025 to February 2026, and found the decline reversed in early 2026: organic CTR on AI Overview SERPs climbed from 1.31% in December 2025 to 2.36% in February 2026 — an 80% increase off the low, against their own model's forecast of continued decline.

Seer also quantified what citation is worth. On informational queries across 2025, average organic CTR was 2.07% when the brand was cited in the AI Overview, 0.94% when it was not, and 3.35% when no AI Overview appeared. Being cited roughly doubles your CTR on an AIO SERP — it does not restore the pre-AIO baseline.

Google's position is that none of this is happening. In August 2025 Liz Reid, VP and Head of Search, claimed total organic click volume was "relatively stable year-over-year" and that Google sends "slightly more quality clicks" than a year prior — defining a quality click as one "where users don't quickly click back" — and dismissed third-party reports as based on "flawed methodologies". Google has published no supporting numbers behind that claim, across three public appearances since. Treat it as a position, not as data.

The scale trap worth avoiding

AI Mode passed one billion monthly active users — and accounted for just 0.34% of all Google searches between January and April 2026, per Similarweb, referring traffic on only 1.6–2.5% of its queries against 17–19% for traditional Google. Both facts are true; most commentary quotes the first and omits the second. The same applies to AI referrals: Similarweb measured ChatGPT referrals growing 52% year over year across ten key industries between September and November 2025, accounting for 87.4% of AI referral traffic in that set — while Conductor put AI platforms at roughly 1% of overall web traffic across major industries in 2025. Growth rates and volume are being conflated constantly.

3. Google told us, in writing, what does not work

In July 2026 Google updated its AI optimization guide with language that settles several arguments at once. On machine-readable files: "You don't need to create new machine readable files, AI text files, markup… Google Search ignores them." John Mueller separately called llms.txt "purely speculative for now", noting the file has existed for years and no AI system uses it. I could find no announcement of support for it from OpenAI, Anthropic, Google, Perplexity or Microsoft.

Schema markup is the more uncomfortable finding, because the correlation is real and the causation is not. Ahrefs ran a matched difference-in-differences experiment published in May 2026: 1,885 pages that added JSON-LD, matched against a control set of 4,000 pages with similar baseline citation levels, measured 30 days either side. The result was no meaningful uplift in citations — AI Overview citations −4.6%, AI Mode +2.4%, ChatGPT +2.2%, the latter two indistinguishable from zero. Across 6 million URLs cited pages were nearly three times more likely to carry JSON-LD, but that correlation is confounded: sites that implement schema also invest in content and earned media. A separate experiment found five AI systems extracted only visible HTML during live fetching, ignoring JSON-LD, Microdata and RDFa entirely.

Google's guide also explicitly rejects content chunking for AI ("there's no requirement to break content into small pieces"), AI-specific rewriting ("AI systems can understand synonyms and general meanings"), pursuing inauthentic brand mentions, and AI-specific schema ("there's no special schema.org markup to add"). Its framing of the whole category: AEO and GEO are still SEO.

4. FAQ rich results are gone

Quietly, with no blog post and no reason given, Google dropped FAQ rich results. They stopped appearing on 7 May 2026; the search appearance filter, rich result report and Rich Results Test support were removed in June 2026; Search Console API support went in August 2026.

This completes a removal that began in August 2023, when FAQ rich results were restricted to government and health sites and HowTo rich results were deprecated on mobile. Across 2025–2026 the direction of travel has been net removal; I could not verify a single new rich result type added in 2026.

Practical instruction: keep FAQ sections where users genuinely have those questions — they serve readers and give fan-out something to match. Drop FAQPage markup added purely for a rich result that no longer exists, and stop inventing questions to justify the schema.

5. Measurement improved — and was deliberately capped

Google launched the Generative AI performance report in Search Console in June 2026, completing global rollout on 31 August. It covers AI Overviews and AI Mode, segmentable by page, country, date and device — and reports impressions only. No clicks, no CTR, no position, no query breakdown. AI Mode clicks are counted, but only inside undifferentiated "Web" totals, where they have sat since June 2025 with no way to isolate them. A logging error from 13 August 2026 produced artificially low impressions in the report's first months; Google confirmed and fixed it within days, but early trend lines deserve scepticism.

Alongside it Google shipped a genuine opt-out, the Search generative AI control, which excludes a property from AI Overviews, AI Mode and Discover AI features without affecting regular Search rankings. Page-level controls are required by March 2027 under UK CMA obligations. One caveat before anyone reaches for that switch: NewzDash measured nearly one in six US trending news queries placing Top Stories inside AI Overviews as of August 2026. Opting out may cost placements you want.

Bing moved earlier and further, adding Intents, Topics, Citation Share and Compare to Webmaster Tools in June 2026, covering Copilot, Bing and select partner AI experiences. It is citations-only, and Microsoft describes Citation Share as "an observational metric, not a ranking system".

That same analysis produced a finding worth acting on. Of 61,332 URLs ChatGPT surfaced as sources, 5,032 became lead sources and only 759 pages were actually opened during thinking mode. An opened page was cited 74% of the time; a page retrieved but never opened, 7%. Whatever makes a page worth opening — clarity of match to the sub-query, extractable answers near the relevant heading — is doing an enormous amount of work.

6. What the evidence says actually drives citation

Strip out the speculation and a short, unglamorous list remains.

Unlinked brand mentions, not backlink counts

Ahrefs ran Spearman rank correlations across 75,000 brands against AI Overview visibility. Branded web mentions came out at 0.664; branded anchors 0.527; branded search volume 0.392; Domain Rating 0.326; backlink count 0.218. Unlinked mentions correlate roughly three times more strongly than raw backlinks. Ahrefs are explicit that this is correlation, that most coefficients are moderate, and that around 26% of brands with zero AIO mentions were excluded — a survivorship filter that likely inflates the numbers. Directionally useful; not a formula.

Updates, not new publishing

Seer studied 7,683 pages carrying 47,097 citations across four brands between March and June 2026. Seventy-five per cent of cited pages had been updated within the last year, 88% within two. Measured by publish date only 42% looked fresh — versus 72% by update date. Gemini is the most freshness-sensitive (78% within a year), Perplexity the most tolerant (65%). For most sites this is the highest-leverage finding in the set: a systematic refresh programme beats a publishing sprint.

Coverage of sub-questions, in extractable prose

Fan-out makes the sub-question the unit of competition. Think of it this way: a page that answers eight related sub-questions clearly, each under its own heading, gives the system eight separate chances to match it. Google has said explicitly you do not need to chunk content for machines — but placing an answer adjacent to the question it answers is simply good structure, and structure is a reasonable guess at what makes a retrieved page worth opening.

Accepting that citation is not attribution

The Semrush and Kevin Indig "Ghost Citations" study measured 3,981 domain appearances across 115 prompts, 14 countries and four engines. Citations appeared in 74.9% of appearances; brand mentions in the answer text in only 38.3%. The engines diverge wildly: Gemini names brands in 83.7% of appearances but cites sources only 21.4% of the time, while ChatGPT cites 87% of the time and mentions brands 20.7%. Measuring "AI visibility" as one number across engines is incoherent — measure per engine, or accept you are measuring noise.

7. Regulation is now moving rankings directly

This is the thread most commercial SEO commentary misses, and for anyone operating across the EU and UK it is the most concrete change of the year. In June 2026 the UK's CMA imposed conduct requirements on Google under Strategic Market Status — a world first. Publishers get effective tools to prevent their content powering AI features; Google must give publisher content proper attribution with clear links in AI-generated results; publishers can separately opt out of fine-tuning use. Google has nine months to implement. The Search generative AI control above is the direct product consequence.

In the EU, the Commission opened a DMA investigation in November 2025 into whether Google demotes publishers under its site reputation abuse policy. On 30 August 2026 Google changed that policy: it removed the specific examples, introduced four holistic evaluation factors, and — critically — manual actions for site reputation abuse no longer affect rankings inside the EEA. Outside the EEA they still do. That is a regulator changing ranking behaviour in one territory and not another, traceable end to end. Separately, the Commission fined Google €890 million in July 2026 for DMA breaches, €460m of it for self-preferencing in Search.

8. And for retail and commerce specifically

The agentic commerce layer is now split between two competing standards and neither has won. Google announced the Universal Commerce Protocol in January 2026 — open source, built with Shopify, Etsy, Wayfair, Target and Walmart, endorsed by Adyen, Mastercard, Stripe, Visa and others. Merchants integrate through Merchant Center product data, keep their business logic and Merchant of Record status, and get one connection serving AI Mode and Gemini. OpenAI's Agentic Commerce Protocol underpins ChatGPT Instant Checkout.

The near-term point for brands: your product data quality in Merchant Center and on retailer PDPs is becoming the input to agentic surfaces, not just to shopping ads. Same discipline retail media teams already run — now doing a second job.

9. What to actually do, in order

1. Audit what AI agents can read. Blocked bots, client-side-rendered copy and thin HTML are the failures I see most often in audits — and any one of them makes everything below pointless.
2. Start a systematic refresh programme. With 75% of cited pages updated within a year, and update date doing the work rather than publish date, a quarterly refresh cycle on top pages beats a publishing sprint.
3. Map sub-questions, not keywords. For each commercial page, list the sub-queries a fan-out would plausibly generate and check you answer each clearly, under its own heading.
4. Invest in mentions, not just links. Being named across credible third-party sources correlates more strongly with AI visibility than backlink volume — PR, partnerships, podcasts, industry press, community presence.
5. Build comparison content. 43.3% brand mention rate versus 18% for informational. Avoiding versus-pages now costs visibility.
6. Set up the reporting you can have — Search Console's generative AI report and Bing's Citation Share — with honest documentation of what they cannot tell you.
7. Stop paying for debunked tactics. llms.txt, AI-specific schema, content chunking for machines, AI-rewritten copy. If an agency line item names one of these, ask for the evidence.

Frequently asked questions

Is GEO actually different from SEO?

Google's own position, published in its AI optimization guide, is that AEO and GEO are "still SEO". The evidence broadly supports that: the tactics that hold up — coverage, freshness, third-party mentions, extractable structure, technical accessibility — are recognisably SEO fundamentals. What genuinely changed is the retrieval mechanism and the reporting, not the discipline.

The label matters commercially, though. If a vendor is selling GEO as a separate service with separate deliverables, ask which of those deliverables has measured evidence behind it.

Should I remove my llms.txt file?

There is no need to remove it — it causes no harm. But Google's documentation states Search ignores machine-readable and AI text files of this kind, no AI vendor has published an announcement of support, and John Mueller has called the concept speculative. Do not maintain it, do not pay for it, and do not treat its presence as AI-readiness.

Was there a Google core update in September 2026?

No. The most recent core update completed on 2 June 2026. August 2026 had a spam update only, which ran 18–20 August. As of 20 September 2026 the gap since the last core update is the longest since 2024 — worth knowing before attributing a traffic change to an update that did not happen.

SE Ranking measured the August spam update's impact across 100,000 keywords: 16.71% of top-10 URLs fell beyond position 100 during 17–22 August, against 9.2% in a comparable July window.

Can I see how much traffic AI Overviews send me?

Not directly. Search Console's generative AI performance report gives impressions for AI Overviews and AI Mode, but no clicks, CTR, position or queries. AI Mode clicks are counted inside your undifferentiated "Web" totals with no way to isolate them. There is no supported method today for attributing sessions to AI Overviews specifically.

Should I block AI crawlers?

The measured pattern among publishers is more precise than "block AI": they block training bots while allowing answering bots, often from the same company. A third-party analysis of Cloudflare Radar data — a single-day snapshot of a sample of top domains, 31 August 2026 — found CCBot, ClaudeBot, GPTBot and Google-Extended all disallowed more often than allowed, while OAI-SearchBot was allowed slightly more often than blocked. Indicative rather than definitive, but the direction is clear.

For most commercial brands the answer is no. You want to be retrievable. The calculus is different for publishers monetising content directly.

Does opting out of AI features protect my traffic?

It protects you from being summarised, not from losing visibility. The Search generative AI control excludes you from AI Overviews, AI Mode and Discover AI features without affecting regular Search rankings — but NewzDash measured nearly one in six US trending news queries placing Top Stories inside AI Overviews. For most brands, opting out removes a surface you would rather be on.

How do I know if AI agents can even read my site?

Check three things: whether your robots.txt permits the answering crawlers, whether your content exists in the initial HTML response rather than only after JavaScript execution, and whether the answer to each question sits close to the heading that asks it. Our AI Score audit runs these checks and returns a prioritised list of what is blocking retrieval.

The honest summary

A great deal changed in the plumbing and very little changed in the work. Retrieval got more diffuse, reporting got thinner, and a year's worth of confident tactical advice turned out to have nothing behind it. What survived contact with controlled testing is coverage, freshness, third-party mentions and clean technical accessibility — the same list as five years ago, now competing for sub-queries you cannot see, measured by reports that deliberately withhold clicks.

The one genuinely new discipline is verification. When the measurement layer is this thin, the difference between agencies is no longer who has the better tactics — it is who checks whether their tactics are real.

Sources

This is a fast-moving area. Every figure above is current as of publication and attributed below.

← All TNOMADS insights