Google AI SEO in 2026: What AI Overviews and AI Mode Actually Change

Google AI SEO is currently the topic with the highest ratio of noise to measurement. Half the industry declares SEO dead; the other half sells a new acronym (GEO, AEO, LLMO, GAIO) attached to the same checklist as before. Neither helps you at 9am when Search Console shows clicks falling while impressions hold steady.

So this guide does three things differently. First, every number has a source and a collection date, so you can verify it yourself — and so you can see that credible studies disagree by a factor of two. Second, what Google actually writes, quoted from the official documentation rather than paraphrased through three blog posts. Third, what we measured on our own site — including a gap this article uncovered in our own setup, which we fixed while writing it.

The short version

• AI Overviews appear on roughly 20% of keywords in Germany (SISTRIX, 100M+ keywords, February 2026). Around 30% in the long tail, considerably less on shopping queries.
• The click loss is real, but the numbers diverge wildly. SISTRIX measures position-1 CTR falling from 27% to 11% in Germany. Ahrefs measures −58% internationally. Both are correct — they measure different things.
• The average hides everything. Across all keywords in Germany, 6.6% of organic clicks are lost. By industry it ranges from 1% to 24%. Your number depends on whether you answer questions or enable transactions.
• Google states no special requirements. From the official docs: “There are no additional technical requirements.” If you are indexed and eligible for a snippet, you are eligible to be cited.
• llms.txt demonstrably does nothing for Google — and Google contradicts itself internally between two teams. Details below.
• The most important thing is measurable, not mystical: can you even see which AI crawler fetched your page? Until we wrote this article, we could not.

AI Overviews vs AI Mode: the technical difference

The two terms get used interchangeably. They are separate products with different consequences for you.

AI Overviews are the AI summary that appears above the classic result list. The ten blue links still exist; they just move down. According to SISTRIX, 78.6% of AI Overviews sit above the organic results, the rest further down the page — Google decides per query what should come first.

AI Mode is a separate conversational surface with no classic result list. In May 2026 Google reported over one billion monthly users, with queries more than doubling every quarter since launch. Since May 2026 it runs on Gemini 3.5 Flash by default.

Both use a technique Google’s own documentation calls query fan-out: instead of searching the user’s question once, the system decomposes it into many sub-questions and runs them in parallel. Google writes:

“While responses are being generated, our advanced models identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links associated with the response than with a classic web search.”

That is the single most consequential technical statement on this topic, and it has a direct implication: you no longer compete for one keyword, but for many sub-questions that were never in your keyword list. An article that answers one main question well and five adjacent ones not at all will not surface for the adjacent ones — even if it ranks first for the main term.

The second implication is genuinely good news: because fan-out gathers more sources than a classic SERP has slots, pages that would never have cracked the top ten for the head term become eligible.

Why AI Overviews do not always appear

Google does not show them universally; it decides per query. They rarely appear on:

• Navigational queries (“YouTube”)
• Simple brand queries
• Many transactional, shopping-oriented searches
• Local queries

They appear frequently on explanation-heavy informational queries combining multiple aspects — health, technical topics, travel planning, purchase research. Which is precisely where most advice blogs live.

The numbers: how far does traffic actually fall?

This is where it gets uncomfortably honest. Published figures range from −15% to −90%, and almost all of them are true in some sense. They measure different things and rarely say so.

What matters here: Ahrefs ran the same study twice with the same methodology — April 2025 produced −34.5%, December 2025 data produced −58%. That is not a contradiction; it is deterioration over time. Ahrefs predicted in the first study that the value was likely the ceiling once novelty wore off. The prediction held.

What usually gets omitted: CTR also fell on keywords without an AI Overview — in Ahrefs’ data from 0.076 to 0.039 between December 2023 and December 2025. Anyone comparing raw before/after wrongly attributes that general decline to AI Overviews. Ahrefs controls for the trend and compares against a forecast instead. That is the methodologically cleaner approach, and the reason the figure is still so large.

The average is the most dangerous number here

SISTRIX puts the total loss in Germany at 265 million organic clicks per month — which sounds catastrophic but represents “only” 6.6% of all organic clicks. That 6.6% is the least useful figure in this article, because it applies to almost nobody:

• Recipe sites: around 1% click loss. An AI summary does not replace a recipe you cook step by step.
• News & media: 7.4%. Time-sensitive and opinion-driven — Google holds back.
• Parenting and baby portals: over 24%.
• Specialised health portals: over 30% (lumedis.de −30.15%, herzstiftung.de −29.66%).
• wetter.com: −0.18%. Booking.com: −0.46%. Amazon: −1.73%.

The pattern is unambiguous and is the only forecast you actually need: the more completely your page answers a question with nothing left to do afterwards, the harder it hits you. Anyone offering a transaction, a tool, a login or a process barely notices.

Largest absolute loser, incidentally: Wikipedia, at 31.6 million clicks per month — which is only 5.56% of its Google traffic. Absolute and relative figures tell different stories.

What Google actually says, verbatim

Most Google AI SEO guides cite other guides. Google’s own documentation on AI features is remarkably short and clear. The central passage on technical requirements:

“To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements.”

Translated: there is no secret format, no AI meta tag, no special schema that lifts you into the answer. If you are indexed and snippet-eligible, you are citable.

Google’s list of practices that remain worthwhile is correspondingly unglamorous:

• Allow crawling in robots.txt — and check your CDN and hosting infrastructure too
• Make content discoverable through internal links
• Provide good page experience
• Make sure important content is available in textual form
• Support text with high-quality images and video
• Ensure structured data matches the visible text

Two of those are systematically underestimated. “And by any CDN or hosting infrastructure” is the most common silent disqualifier: an aggressive bot protection layer renders the best content strategy irrelevant, and no SEO tool will show you this. And “available in textual form” hits every site that delivers its key claims only inside graphics, video, or client-side rendered components.

The control nobody mentions

If you want to stay out of AI answers, Google gives you exactly these levers: nosnippet, data-nosnippet, max-snippet and noindex. There is no switch that disables AI Overviews while leaving normal search untouched — Google justifies this by saying AI is integral to how Search functions.

data-nosnippet is the most precise instrument: it excludes individual sections of a page from snippet use without touching indexing. For sites with expensive original research, it is often the only sensible middle ground between “give everything away” and “become invisible”.

What demonstrably does nothing: the llms.txt case

Few topics illustrate how consulting fashions form quite so well. llms.txt is a file at the domain root meant to explain a site’s content to language models in structured form. Proposed in 2024, sold as mandatory ever since — and the state of evidence is unambiguous:

• Google’s official AI optimisation guide has, since 15 May 2026, explicitly listed machine-readable files like llms.txt among the tactics you can ignore for AI Overviews and AI Mode citations.
• John Mueller publicly compared it to the long-discredited keywords meta tag.
• There is no signal that it influences inclusion or ranking in Google’s AI surfaces.

And here is the part rarely told: Google contradicts itself. While the Search team waves it off, Lighthouse 13.3 checks for llms.txt by default and flags sites that lack it. Two teams inside one company, two positions, no resolution.

The resolution is not a matter of opinion but of scope: Search Central talks about visibility in search; Lighthouse talks about agentic browsing — whether an AI agent can complete a task on your site. Different problems. Mueller’s framing: markdown versions are useful for documentation, not for most websites, and you should “prioritize needs before dreams”.

Our position: getmind.io has no llms.txt, and we are deliberately not adding one while there is no measurable effect. If we do add one, we will write here what measurably changed. A green tick in an audit tool is not a result.

What actually helps: content that can be quoted

If there are no special technical requirements, the shape of the content decides. SISTRIX’s analysis of frequently cited pages and Google’s own fan-out description point in the same direction.

1. Every section must stand on its own

Fan-out retrieves passages, not pages. A paragraph that makes no sense without the three before it cannot be quoted. Concretely:

• Headings containing the question, not just the topic (“How much does X cost?” instead of “Costs”)
• The answer immediately below the heading, not after two paragraphs of throat-clearing
• Avoid pronouns referring back to earlier sections
• Comparison tables with clear column headers — tables are extremely quote-friendly

SISTRIX finds numbered listicles, step-by-step instructions, comparison tables, and FAQ blocks with cleanly separated question-answer units perform well. What loses: general glossary text with no concrete utility.

2. Make provenance and freshness explicit

AI systems favour sources with traceable origins:

• JSON-LD with publisher and dateModified
• Visible author information with recognisable subject expertise
• A visible “last updated on …” in the page body, not only in markup

Domains with a clear topical focus are cited more often than generalists. Specialisation beats breadth — more sharply in AI answers than in classic search.

3. Where most people fail: original contribution

Google’s documentation and SISTRIX’s analysis say the same thing in different words: fully AI-generated summaries of existing content are not needed in AI search systems — those systems can summarise perfectly well themselves. What gets cited is what the model cannot generate: your own measurements, your own failure logs, named figures with collection dates, concrete prices, real screenshots, a documented incident.

That is not a moral argument but an economic one: if your text is the summary, then the summary is your competitor.

Measurement: the part where we caught ourselves out

Every article on this topic recommends checking your server logs for AI crawlers. We wanted to do exactly that with real getmind.io numbers — and discovered we could not. Our web server wrote no access log for getmind.io at all. No log block in the configuration, no file, nothing.

That is the most uncomfortable way to test a recommendation: you set out to demonstrate it and realise you never implemented it yourself. It never surfaced because nothing is missing when a log is missing — there is no error, no empty view, no red status. Only a question you never ask, because you suspect there is no answer.

We set it up while writing. And in doing so, made a second, more instructive mistake.

caddy validate said “Valid configuration” — and the config would not load

After adding the log block, we validated properly:

caddy validate --config /etc/caddy/Caddyfile --adapter caddyfile
# → Valid configuration

Green. Then the reload — which ran into a timeout until systemd killed it after 90 seconds. The journal held the actual reason:

loading config: setting up custom log 'log0': opening log writer using
&logging.FileWriter{Filename:"/var/log/caddy/getmind-access.log", ...}:
open /var/log/caddy/getmind-access.log: permission denied

The log file had been created by root; the service runs as user caddy. validate checks syntax, not reality — it never asks whether the process may actually create the file it is told to create.

The most interesting part came next. The website kept serving HTTP 200 throughout. Not because everything was fine, but because the reload had failed and the old configuration kept running. A curl against the homepage would have reported “all good” while the change had in fact never landed.

Two rules from this, both wider than this incident:

1. A “valid” result is a statement about form, not about executability. After every reload, check the service status, not the website.
2. A 200 after a failed change only proves the old state is still running. Verify the effect of the change, not the reachability of the service.

After chown caddy:caddy and a clean restart, the log writes. The first entry was our own test request — and from now on we can answer the question this section raised. What we still cannot do: present numbers from it. A log that started this morning has no history. Anything else would be invented.

What Search Console actually shows you

Google counts clicks and impressions from AI Overviews and AI Mode inside the normal search data, under the “Web” search type. There is no separate filter and no dedicated report. That means two things:

• You cannot read off the AI share of your traffic directly.
• The fingerprint is indirect: impressions stable or rising, position stable, clicks falling. If you see that, an AI answer is probably sitting above you.

Alisa Scharf of Seer put it well in May 2026: “Can’t optimize what you can’t measure” — coupled with the observation that Bing Webmaster Tools delivers considerably more here than Google’s free tooling. That is another reason not to ignore Bing SEO: not for Bing’s market share, but for the data.

One bright spot from Google’s documentation, in fairness:

“We’ve seen that when people click from search results pages with AI Overviews, these clicks are higher quality (meaning, users are more likely to spend more time on the site).”

That is Google’s own, third-partyly unverifiable claim — but it matches the logic: someone who clicks despite having an answer has a real reason to. So measure dwell time and conversions alongside clicks. Fewer but better-qualified visitors is a different outcome from simply fewer visitors.

How user behaviour is changing — and what it means for keywords

Google’s own AI Mode data (May 2026, US) is more revealing than the headline user count:

• AI Mode queries are three times longer than classic searches
• Follow-up queries increase 40% month over month
• Over 16% of searches are multimodal (voice, image, video)
• Planning queries grow at 80% of the overall rate

Jeffrey Cohen (Skai) framed it better than any analysis:

“Shoppers aren’t typing ‘running shoes.’ They’re asking ‘what are the best running shoes for a wide foot that I can wear for a half marathon training on pavement under $150.’ That’s not a keyword. That’s a brief.”

The practical consequence: search volume alone no longer works as a keyword evaluation metric. Whether an AI Overview triggers, and how hard it suppresses CTR, belongs in every keyword analysis. Two keywords with identical volume can differ threefold in actual traffic. In practice: use the SERP feature filters (both SISTRIX and Ahrefs have them) and score AI Overview keywords differently from the rest.

One structural point too: when users ask questions in briefing form, pages that cover constraints win — price ceilings, edge cases, exclusion criteria, “who this is not worth it for”. Exactly the sections classic SEO copy omits because they do not serve the head term.

The honest counter-argument: where this topic is overstated

So this article does not become its own kind of panic, three objections to its own thesis:

1. For many websites, almost nothing changes. Google rarely serves AI Overviews on transactional queries. If you run a shop, a SaaS product behind a login, or a local service, your click loss is probably closer to Amazon’s 1.73% than to health portals’ 30%. Before rebuilding your content strategy: measure your own number instead of adopting an industry average.

2. Most “GEO measures” are classic SEO with a new name. Clean structure, clear headings, structured data, fast pages, genuine expertise — all of that was in every decent guide already. Paying an agency that bills a new acronym for it is paying for a rename.

3. The field moves faster than optimisations take effect. Eight months separated Ahrefs’ two measurements, and the figure worsened from −34.5% to −58%. Building a strategy on one percentage today means building on a snapshot. Only the direction is reliable: more answers without clicks, more weight on brand and recognition, less weight on ranking alone.

The most honest line we found came from Jake Ward:

“Search is very much alive, just different. We’re moving further into a world of visibility > clicks.”

With one caveat worth stating: visibility without clicks is hard to bank. If your living depends on people reaching your site, do not accept “visibility” as a consolation prize — treat it as an intermediate goal that must eventually convert into something measurable.

A concrete checklist

Sorted by effort, not by volume of hype:

Immediate and free:

1. Check whether you have access logs at all. If not, set them up — without them, every statement about AI crawlers is a guess. (We had to do exactly this while writing.)
2. Check robots.txt and your CDN/host bot protection. Google names this explicitly; it is the most common silent exclusion.
3. In Search Console, filter for pages with stable position, stable impressions and falling clicks. That is your exposure list.
4. Measure dwell time and conversions on those pages, not just clicks.

Moderate effort:

5. Rewrite headings as questions and put the answer directly underneath.
6. Add FAQ sections with cleanly separated units — ideally with FAQPage markup whose text matches the visible text.
7. Maintain dateModified and show a visible update date.
8. Add comparison tables where they genuinely fit.

Higher effort, highest impact:

9. Put in something a model cannot generate: your own measurements, prices with dates, a documented failure, real screenshots.
10. Cover the constraints: “who this is not worth it for”, exceptions, cost ceilings.
11. Focus topically instead of spreading wide.

What you can skip: llms.txt for Google, AI meta tags, keyword-density optimisation for language models, and any offer that “guarantees placement in AI Overviews”. There is none.

How we handle this on this site

So this does not stay at the level of advice, here is getmind.io’s status as this article goes live:

• Access log: in place since today (see above). Before that: none.
• Article JSON-LD with datePublished, dateModified, author, publisher: present, verified live.
• Anchorable headings: yes, every h2/h3 has an id.
• Content in the served HTML without JavaScript: yes — this article is fully present in the source.
• FAQPage markup on blog articles: no. Our tool pages have it, our blog articles do not — even though nearly all of them have an FAQ section. That is an open gap on our side and sits as a to-do in our editorial list. We are writing it here rather than hiding it.
• llms.txt: no, deliberately.
• Visible update date: no, only in markup. Also open.

Three of seven items are not done on our own site. An article that presents a checklist while implying its author has completed all of it is, in most cases, not being straight with you.

FAQ: Google AI SEO

What is Google AI SEO?

Google AI SEO means optimising content to be cited and linked as a source inside Google’s AI features — AI Overviews and AI Mode. Technically, Google states there are no additional requirements beyond normal SEO: the page must be indexed and eligible for a snippet. The difference lies in the shape of the content, not in the technology.

How much traffic do AI Overviews actually cost?

It depends heavily on your sector. The German average is 6.6% of organic clicks lost (SISTRIX, February 2026), but the range runs from about 1% for recipe sites to over 30% for specialised health portals. On position 1, CTR falls from 27% to 11% on keywords with an AI Overview; Ahrefs measures −58% internationally. Do not rely on the average — measure your own figure in Search Console.

On how many searches do AI Overviews appear?

On roughly 20% of keywords in Germany, a level that has been broadly stable for months (SISTRIX, over 100 million keywords). In the long tail the rate is around 30%, and considerably lower on shopping-related queries. 78.6% of AI Overviews are displayed above the organic results.

What is the difference between AI Overviews and AI Mode?

AI Overviews are AI summaries inside the classic search results page — the ten blue links remain but move down. AI Mode is a separate conversational surface with no classic result list, with over one billion monthly users (Google, May 2026). Both use query fan-out, but they may use different models and therefore surface different sources.

Do I need an llms.txt file for Google?

No. Google’s official AI optimisation guide has, since May 2026, explicitly named machine-readable files like llms.txt as a tactic you can ignore for AI Overviews and AI Mode; John Mueller compared it to the discredited keywords meta tag. The only confusing part is that Lighthouse 13.3 checks for it by default — but that concerns agentic browsing, not search visibility.

Can I see how much traffic comes from AI Overviews in Search Console?

Not directly. Google counts clicks and impressions from AI Overviews and AI Mode inside the normal search data under the “Web” search type, with no separate filter. The indirect fingerprint is stable impressions and stable position combined with falling clicks.

How do I stop my content appearing in AI Overviews?

Through nosnippet, data-nosnippet, max-snippet or noindex. There is no switch that disables only the AI features while leaving normal search untouched — Google justifies this by saying AI is integral to Search. data-nosnippet is the most finely dosed option, since it excludes individual page sections without affecting indexing.

Is SEO dead because of AI search?

No, but the measurement of success is shifting. Google builds its AI features on the same search index — if you are not indexed and not judged a relevant source, you cannot appear in any AI answer. Classic SEO therefore remains the prerequisite. What changes: ranking alone no longer guarantees clicks, and content that a summary already fully replaces loses its basis.

Conclusion

Google AI SEO is less a new discipline than a shift in how success is measured. The technical prerequisites are, per Google, unchanged — indexable, snippet-eligible, cleanly structured. What changes is the yield: the same position returns fewer clicks, and how many fewer depends almost entirely on whether your page conclusively answers a question or enables something an answer cannot replace.

Two honest closing points. First: the industry average is worthless to you — the gap between 1% and 30% click loss is the gap between “ignore this” and “rethink the business model”, and only your own Search Console produces that figure. Second: check whether you can measure at all before you optimise. While writing this article we discovered we had no access log — and we only noticed because we set out to demonstrate our own recommendation. Missing measurement never announces itself.

If you want to keep going: how to become visible in the second major AI data source is covered in our guide to Bing SEO — where, incidentally, the free Webmaster Tools deliver exactly the AI data Google does not (yet) show. Which tool reliably surfaces SERP features and AI Overviews is settled in our SISTRIX vs Ahrefs comparison. If you work on WordPress, the technical groundwork is in our overview of WordPress SEO plugins, and if you want to check load times and page experience, you can run our Core Web Vitals Test right here.