The same article comes out every autumn. It's called "15 SEO Trends for Next Year," and inside you'll find AI, voice search, E-E-A-T and video. There's nothing to verify: no dates, no samples, no link to wherever the number came from.
We took a different route. What follows is only what you can open and read for yourself: Google's update log, the Search Central documentation, official blogs from Google and Microsoft, and studies with disclosed methodology. Where the data doesn't exist, we say so. The year isn't over, so everything concerning 2026 is counted through 9 October inclusive.
One warning up front, and it matters more than anything else in this text. The figures below were gathered mostly from English-language search results in the US and Europe. There is no publicly available research on iGaming, betting or the gambling vertical with sound methodology. Not one study. So apply these percentages to your own networks with your brain switched on — and we'll come back to this point at the end. The underlying mechanics all of this sits on we covered separately, in our guide to SEO for affiliate marketing (in Russian); this piece is only about what changed over the year.
What actually changed in search during 2026
Let's start with the updates, since that's usually where the conversation stops.
In less than a year, Google logged seven ranking updates in the Search Status Dashboard: a Discover update in February (5 February, 21 days 17 hours), a spam update on 24 March (19.5 hours), a core update on 27 March (12 days 4 hours), a core update on 21 May (11 days 21 hours), and three more spam updates in June, August and September. The September one started on the 24th and kept rolling until 8 October — almost two weeks.
Look at the ratio. Only two core updates in nine months; everything else was spam work. For verticals used to counting updates one by one and waiting for "the big March one," the picture looks different: Google spent nine months mostly cleaning the index rather than broadly reassessing quality. What exactly it cleaned, the company won't say, as usual, and to questions about drops it gives the standard answer — a decline doesn't imply a violation, and there's no specific fix.
The second change got far less attention, even though it changes day-to-day work more than any update.
On 3 June, Google shipped Search Generative AI performance reports in Search Console. For the first time you can see how often your pages surface in AI Overviews, AI Mode and the AI features in Discover, broken down by page, country, device and date. A note was later added to the same Search Central post: as of 31 August, the reports are available to all sites worldwide.
Now two caveats that keep the report from solving the problem and leave it merely describing one. There are no clicks, only impressions. And those impressions are already counted in the main Performance report — so adding the two together to get "total visibility" won't work, you'll be double-counting. Alongside this, Google added a toggle that lets you remove a site from generative features entirely.
Microsoft got there first: on 10 February, Bing Webmaster Tools gained an AI Performance report in public preview. It shows citations, cited pages and grounding queries — the phrasings the AI used when it came looking for your content. No clicks there either, but there is a query-level breakdown, which Google doesn't have. For reconnaissance that's more useful than it sounds, and it takes five minutes to connect.
The third change is quiet, and unpleasant for anyone who lived off markup. Since 7 May, FAQ rich results no longer appear in search results. In June, Google removed the corresponding search appearance and report in Search Console along with support in the Rich Results Test; in August, support in the Search Console API. This wasn't announced with a blog post but with a notice in the documentation. You can leave the markup in place: Google says other systems are free to keep processing it.
And one more. At I/O on 19 May, Google stated that AI Mode has passed one billion monthly users and that queries within it have been more than doubling every quarter since launch. Gemini 3.5 Flash became the default model there. These figures can't be verified from outside — it's a corporate statement, and it should be treated as one.
How many clicks AI actually takes
Here 2026 delivered something the previous years didn't: experiments rather than observations.
First, the most-cited work. On 4 February, Ahrefs took 300,000 keywords — half with AI Overviews, half informational without them — pulled average desktop CTR from aggregated Search Console data, and compared December 2023 with December 2025. Position-one CTR on keywords with an AI block fell from 7.3% to 1.6%. The modelled contribution of AI Overviews themselves: minus 58% for position one. For positions two through ten, between minus 50.8% and minus 19.4%.
And now the number almost nobody carries over from that study. Informational keywords without AI Overviews lost 49% of position-one CTR over the same two years, from 7.6% to 3.9%. Clicks aren't only draining into AI answers. The results page keeps filling with blocks regardless, and pinning the entire decline on one cause is exactly the mistake that comes back to you from a client or an affiliate programme.
There's also the method: Ahrefs models the effect on observational data, it doesn't prove causation. Others did.
In January and February, Saharsh Agarwal and Ananya Sen ran a randomised field experiment: 1,065 people in the US, a Chrome extension, three groups — ordinary Google, Google with AI Overviews forcibly stripped out, and forced AI Mode — two weeks per participant. AI Overviews showed up on 42% of queries. With them removed, outbound clicks rose to 0.61 per search from 0.38. On affected queries, organic lost 38% of clicks, and the share of searches ending with no click at all rose from 54% to 72%.
One separate detail that undercuts the "users prefer it this way" argument: self-reported search satisfaction among people without AI answers barely changed.
On 18 August came a second preregistered experiment — Wang, Gleason, Bart, Wilson and Metaxa, 1,100 participants. Directionally the same conclusion: remove AI Overviews and AI Mode and referrals to sites go up; push a person into AI Mode alone and referrals fall while trust in the information found declines. An honest caveat from the authors: midway through the experiment Google changed the HTML structure of AI Overviews, after which the hiding mechanism worked in roughly half of cases. This is a preprint and has not been peer-reviewed.
For context, the Pew study on March 2025 data: a panel of 900 people; with an AI Overview present, users clicked a regular link in 8% of visits versus 15% without one, and clicked links inside the block itself about 1% of the time. Google publicly called the methodology flawed at the time. The company still doesn't publish its own click data from AI Overviews.
What this means for the work. Clicks fall when an AI answer is present. That's established; there's nothing left to argue about.
What it doesn't mean. That you can plug minus 38% or minus 58% into your own model. All of these studies are US-based, English-language results, mostly desktop, with differing query mixes. That's precisely why estimates of AI Overview frequency in 2026 publications range from 8% to nearly 50% — it's panel variation, not different states of Google. For gambling, betting and grey verticals there is no data of this quality at all, and anyone quoting you a precise percentage for iGaming made that percentage up.
Briefly on "AI traffic replacing organic." Similarweb's panel data shows referral traffic from generative platforms is still orders of magnitude smaller than Google's. Absolute figures differ several-fold between vendors, so it's only honest to speak of an order of magnitude. No substitution is visible in the data.
GEO, AEO and the other acronyms: what Google itself said
In 2026, Google finally answered with documentation instead of employee tweets.
In its guide to optimising for generative AI features, last revised on 10 July, it states plainly: from Google Search's perspective, optimising for generative AI search is optimising for the search experience, and thus still SEO. Generative features run on top of the core ranking systems via RAG and query fan-out — the model pulls pages from the index while generating additional related queries in parallel.
Then comes a section worth printing out and pinning above your contractor's desk. Google says that:
- llms.txt and other "special" files are not used by Google Search — no benefit, no harm;
- there's no need to chop content into small chunks, and there's no ideal page length;
- there's no need to rewrite text specifically for AI, since the systems understand synonyms and meaning;
- structured data isn't required for generative features and there's no special schema for them, though markup is worth keeping for ordinary rich results;
- chasing inauthentic mentions is less useful than commonly assumed.
Google separately warns that churning out pages for every possible query variant, fan-out phrasings included, in order to influence results is scaled content abuse. Which is exactly what half the "GEO content" generators currently on sale are built for.
The technical minimum for eligibility is spelled out clearly: the page must be indexed, eligible to be shown with a snippet, and not excluded from generative features in Search Console. Eligibility doesn't guarantee being shown.
Now for evidence from the other side. We found no independent controlled studies where a specific GEO technique causally raises visibility in AI answers. The largest piece of work this year is correlational: Ahrefs ran 75,000 brands through millions of AI responses and found the strongest relationship with YouTube mentions (around 0.737) and branded web mentions (0.66–0.71). Ahrefs itself notes these are correlations and that inflating the metrics isn't obliged to lift visibility. Along the way it emerged that ChatGPT, AI Mode and AI Overviews largely mention the same brands — an overlap of 0.779.
Formally, Google and Ahrefs don't contradict each other: Google criticises artificial mentions, Ahrefs measures all of them. In practice, though, people draw opposite conclusions from the two, so when someone hands you a deck titled "mentions are the new backlinks," it's worth asking which source it rests on.
Content: what Google wrote about quality in plain words
The most concrete formulation of the year is the distinction between commodity and non-commodity content in that same guide.
Commodity content, per Google, is material based on common knowledge that anyone could have written. Their example: "7 Tips for First-Time Homebuyers." Non-commodity content carries expert or lived experience beyond common knowledge. Their example: an account of why the author waived the inspection when buying a house and how that turned out.
The criterion Google names as decisive for long-run visibility in generative features is a unique point of view. A first-hand review offers a perspective a summary doesn't have. And as a separate point: don't recycle what's already been said online or what a language model could easily produce.
The position on AI in content production hasn't changed and remains procedural. Using the tools is fine; the output must meet the Search Essentials and the spam policies — above all, that same scaled content abuse policy.
What the documentation does not contain is E-E-A-T as a ranking factor. It remains a set of quality assessment criteria, not a lever you can turn.
Technical work and measurement
Two technical topics carry into 2027 with documented grounds behind them rather than blogger consensus.
The first is crawl accessibility as a condition of eligibility. Google explicitly ties visibility in generative features to indexing, snippet eligibility and content being reachable by the crawler, including ordinary JavaScript SEO and crawl budget management on large, frequently updated sites. On semantic HTML its position is relaxed: useful, but chasing perfect validity isn't the point.
The second is that rich results have stopped being a dependable channel. The FAQ story is the model case: restricted in 2023, switched off entirely on 7 May 2026, with the reporting dismantled right after. Any scheme that lives on extra SERP real estate earned through markup should now be treated as unstable by default. No transition period, no blog post, just a note in the docs.
Measurement is the more interesting part. In 2026 two official AI visibility metrics appeared, one from Google and one from Microsoft. Both show presence; neither shows clicks. Google's report additionally offers no query breakdown and no average position.
Hence three mistakes that will be everywhere over the next year:
Adding impressions from the generative features report to impressions in the main Performance report. They're already in there.
Comparing the incomparable. Ahrefs' minus 58% is a modelled desktop effect from aggregated GSC data, December 2023 against December 2025. The experiment's minus 38% is a causal effect on a US panel in January and February 2026. These are different quantities produced by different methods, and putting them on one slide as a "trend" is wrong.
Dumping four things into one metric: search impressions, impressions in AI answers, brand mentions, visits and conversions. Four distinct entities from four distinct sources.
Five forecasts for 2027
AI reporting will expand, but it won't become complete. Grounds: the launch of the reports in Search Console on 3 June and in Bing Webmaster Tools on 10 February, and Google completing its rollout on 31 August. Google wrote in its announcement that metrics would be added over time, and Microsoft called its report a preview outright. What this does not mean: neither company promised clicks from AI answers, and neither named a date. Don't plan around them.
Position one will keep losing its link to traffic volume. Grounds: three independent measurements — Ahrefs and two randomised experiments. Practical consequence: forecasting traffic from rankings makes less sense every quarter; the maths has to be done by query segment, separating where the AI block fires from where it doesn't. Limitation: the size of the effect depends on topic, language, country and device, and for our verticals there are no confirmed figures.
No separate "GEO algorithm" will appear in Google's documentation. Grounds: the 10 July guide, which describes generative features as a layer over core ranking and lists the non-working tactics by name. So the working setup for 2027 stays as it was: proper SEO plus a separate measurement of presence in AI answers. Limitation: this is Google Search's position. It doesn't extend to ChatGPT or Perplexity, and there's simply no public data of comparable quality on how those work.
Rich results will keep shrinking. Grounds: the full FAQ cycle — restricted in 2023, switched off in May 2026, reporting cleared out by August. Conclusion: don't build traffic into your model if it depends on one specific rich result type. Limitation: which types go next, or whether any do, can't be predicted — Google doesn't publish its plans.
Agentic interfaces will stay a watching brief, not an investment. Grounds: the I/O announcements — information agents, generative UI, the expansion of Personal Intelligence. Information agents went first to AI Pro and Ultra subscribers only, and in its documentation Google frames agent recommendations as optional, in the spirit of "if this is relevant to your business and you have the spare time." Limitation: there is no data whatsoever on how agentic scenarios affect site traffic. Any figure on the subject today is invention.
What to do right now
Immediately. Split reporting into two tracks: clicks and positions on one side, impressions in generative features on the other, never summed. Rebuild traffic forecasts by query segment instead of average position. Export the history for any search appearance whose reporting is being removed — as FAQ showed, the export window is short. Check the technical minimum for eligibility: indexed, snippet-eligible, and make sure nobody flipped the Search Console toggle by accident. Connect Bing Webmaster Tools; it has grounding queries, which Google doesn't provide.
Test and measure. Brand presence in AI answers as its own metric, kept separate from visits. The mentions hypothesis, YouTube included: the Ahrefs data gives you grounds to test it, not grounds to believe it. The link between appearing in an AI answer and a lift in branded demand — there's no public data here, so your own measurement will beat anyone else's estimate.
Don't rush. Into llms.txt for Google's sake: it ignores the file, stated in plain text. Into rewriting content for AI and slicing it into chunks: Google calls this unnecessary. Into generating pages for fan-out queries: that's explicitly classed as scaled content abuse. Into buying mentions for AI visibility: a correlation in the Ahrefs data doesn't mean purchased mentions will work. And into transplanting American percentages onto your own projects.
In short
2026 didn't bury SEO, and it didn't confirm a single one of the scenarios being sold from conference stages.
It did three things. It moved the argument about clicks from observation to experiment — and the experiments confirmed the decline. It produced the first official, if truncated, AI visibility metrics: Microsoft in February, Google in June. And it put Google's position on GEO and AEO into documentation, declaring a sizeable share of the tactics currently on sale useless.
The main shift heading into next year isn't in methods, it's in metrics. The click has stopped being the only currency, and what replaced it still shows half the picture.
The most honest thing to say about forecasts is that their accuracy is bounded by the data available. For English-language US results, there's plenty. For iGaming, affiliate and Russian-language search, there is none — and any confident forecast for those segments today rests not on evidence but on somebody's wishful thinking.
