The state of AI search in 2026: what the data actually says

In the Carnegie Mellon and Indian School of Business field experiment, AI Overviews appeared on 42% of the Google searches participants ran. How often they show varies by query type, and published trackers disagree on the exact rates, so check your own priority queries by hand instead of trusting one headline number.

That is the cleanest number in the space, because it came from a real randomized experiment, not a correlation. When the researchers hid AI Overviews for one group, outbound clicks rose from 0.38 to 0.61 per search, which means AI Overviews cut organic clicks by about 38% on the searches where they appeared. The same study found user satisfaction was unchanged whether the AI answer was shown or hidden, which directly contradicts the “people are happier with less clicking” line.

An Ahrefs study of 300,000 keywords found the click-through rate on the number-one organic result fell about 58% on keywords that trigger an AI Overview. The click did not move down the page. It disappeared. This is why “we still rank #1” is no longer the same as “we still get the traffic.”

For twenty years, ranking in Google’s top 10 was the whole job. In 2026 it is the entry fee for one game and close to irrelevant for the other. Google’s own documentation explains why. Google says AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related searches across subtopics and data sources to build a response, and that the two features may use different models and techniques, so the links they show will vary. A page has to be indexed and eligible for a snippet to be cited at all, but ranking for the head query does not decide whether it gets picked.

Read that again. The same company, answering the same question two slightly different ways, can show different sources. Optimizing for one surface does not guarantee the other, and optimizing for classic blue-link rank guarantees neither.

None of this means SEO is dead. The technical foundation that makes you rank, crawlable HTML, fast pages, clean structure, real content, is also what makes you eligible to be cited. It is necessary. It is just no longer sufficient, and treating it as the finish line is the mistake.

This is the finding that reframes the whole discipline. Muck Rack analyzed 25 million links across the major engines and found about 84% of AI citations come from earned media, journalism, research, and reference sources, not from a brand’s own website. AI systems treat third-party editorial coverage as the proxy for whether a brand is real and worth naming.

So the old link-building reflex is weak here. A backlink helps, but an unlinked mention in a publication an engine already trusts can help more, because the engine learned the brand from that source in the first place. The practical translation: the highest-value work is getting referenced in the places AI reads, then making sure your own site is clean and current enough to be the page it links when it does name you. We go deeper on this in our breakdown of what a real AEO agency actually does.

Treating “AI search” as one thing is a mistake. The major engines pull from visibly different places, and the 2026 source studies make the differences usable:

• ChatGPT cites sources in about 96% of its responses but averages only about five citations each, and its most-cited domain is Wikipedia, according to Muck Rack’s 25-million-link analysis.
• Perplexity was not part of that Muck Rack sample, so test it directly: run your category questions and log which publications and communities it cites.
• Claude cites in only about 55% of responses in the same Muck Rack data, but averages 13 sources when it does, and leans on research sources such as PubMed Central.
• Google AI Overviews names some of the same tools as ChatGPT while reading almost none of the same pages. DerivateX ran 15 buyer-intent queries through both engines and logged 402 citations in one SaaS category: on the open-ended queries the two recommended the same tools about 32% of the time, yet cited the same sources only about 4% of the time. AI Overviews also pulled wider and more vendor-owned (16.3 sources per answer against ChatGPT’s 10.5), so a citation from one engine tells you very little about the other.
• Gemini leans on Reddit more than the other engines in the Muck Rack data. Where your category gets discussed in public communities matters, not just what your own site says.

The takeaways are concrete. Keep key pages current and dated, so an engine looking for a fresh answer has a reason to pick yours. Put real, sourced statistics on the page: in the original GEO study, changes such as adding statistics, quotations and source citations raised visibility in generative engine answers by up to 40%. And do not ignore where your category actually gets discussed, because the engines read those places too.

Being the frontier means saying the unglamorous part out loud. Several tactics that got sold hard in 2024 and 2025 do not hold up against 2026 evidence:

• llms.txt as a citation signal. Google’s guidance on AI features says you do not need to create new machine-readable files or AI text files to appear in AI Overviews or AI Mode. It is fine as hygiene and genuinely useful for agentic and developer tools, but do not count on it to get you cited in ChatGPT, Perplexity, or AI Overviews. Read our honest take in the llms.txt explainer.
• High rank as a citation guarantee. Covered above: the correlation collapsed.
• Domain authority as the predictor. It barely correlates with AI citation. Topical authority and E-E-A-T do.
• Coordinated mentions and content chunking. Google explicitly flagged these as not working, and in some cases counterproductive.

One more honest note on the platforms: Google now publicly calls AEO and GEO “still SEO” and resists the idea that it is a separate discipline. Microsoft went the other way, formalizing GEO in its guidelines and shipping the first real first-party AI-citation dashboard inside Bing Webmaster Tools. The vocabulary is still being argued over. The underlying work, structure, authority, and earned coverage, is not.

If the on-page numbers feel abstract, the buyer-behavior numbers are not. Forrester’s 2026 survey of about 18,000 business buyers found twice as many named generative AI as their most meaningful research source than any other option, ahead of vendor websites and sales reps. 94% reported using AI somewhere in the purchase. More than half now build an internal shortlist before they ever touch a vendor’s site. Forrester’s own summary is blunt: the model of driving traffic to your site to then nurture them is getting much less effective, because the research happened somewhere you were not invited.

Local is where this gets sharp for most Calgary businesses. The SOCi Local Visibility Index found only about 1.2% of business locations get recommended in ChatGPT, against 35.9% that show up in Google’s local 3-pack. AI is dramatically more selective about which local businesses it will name. Worse, when AI does describe a local business, it matches the Google Business Profile data only about 68% of the time. It is confidently wrong about a third of the time. And about 45% of consumers now use AI for local business recommendations.

We have local data of our own. When we ran 50 ChatGPT queries about Calgary marketing agencies, nine firms received nearly all the citations, and three signals predicted who got named almost perfectly. And when we audited 152 Calgary small business websites for AI readiness in our State of AEO in Calgary study, zero scored Strong. The demand is here, the engines are answering, and almost nobody local is set up to be the answer. That gap is the opportunity, and it closes as competitors wake up.

Strip it down and the playbook is short:

• Fix the foundation. Fast site, clean and accurate schema, current content with real dates, and crawler access for the AI retrieval bots. This is table stakes and most sites still fail it.
• Earn third-party coverage. This is where 84% of citations come from. Get referenced in the publications, communities, and reference sources your category’s engines actually read.
• Keep it fresh and quotable. Update key pages, and give them specific, sourced numbers worth pulling. Statistics and quotations are among the few on-page changes with published test results behind them.
• Win your entity, especially locally. Consistent, accurate listings and profiles, because AI is selective and frequently wrong about local businesses.
• Measure it. Google Search Console still shows nothing about AI citations. Track presence across the engines directly, with the AI search visibility tools built for it.

The brands that treat this as a passing trend will spend 2026 watching their traffic erode without a clear cause. The ones that treat it as the new shape of search will be the names the answer gives. In a city where almost no one is ready, being early is the entire advantage.

1. Carnegie Mellon & Indian School of Business, randomized field experiment on AI Overviews and clicks, via Search Engine Journal, April 2026.
2. Ahrefs, 300,000-keyword CTR study, via MediaNama, February 2026.
3. Seer Interactive, AI Overview CTR longitudinal study, April 2026.
4. Muck Rack, “What is AI Reading?” 25M-link citation analysis, May 2026.
5. Forrester 2026 B2B Buyers’ Journey Survey (~18,000 buyers), January 2026.
6. SOCi Local Visibility Index 2026 and BrightLocal Local Consumer Review Survey 2026.
7. Google AI search guidance, May 2026; Microsoft Bing Webmaster Tools AI Performance, February 2026.
8. DerivateX Research, “The Agreement Gap” ChatGPT vs Google AI Overviews benchmark (15 buyer-intent queries, 402 citations), June 2026.