Is SEO Dead Because of AI? An Evidence Check

What would have to be true for SEO to actually be dead?

SEO would be dead if the systems people use to find businesses had stopped reading web pages to build their answers. Not one of them has. Google AI Overviews, Google AI Mode, ChatGPT Search, Perplexity and Gemini all retrieve live documents and compose over them, so the indexed web is still the raw material and a page that cannot be found, parsed or quoted still cannot appear. That is the whole discipline, intact by definition.

So the literal question has a boring answer. The one underneath it is sharper: will the work still pay me the way it used to? There the answer is no, and pretending otherwise is how agencies lose clients. Three things changed measurably between 2024 and 2026, all of them published, dated, and checkable against your own analytics this week.

Change one: how much did zero-click search actually grow?

From 60.45% of United States Google searches in 2024 to 68.01% across January through April 2026, according to SparkToro's analysis of Similarweb clickstream panel data, published by Rand Fishkin in June 2026. The same metric sat near 49% in 2019, and Fishkin calls the 2024 to 2026 move the fastest acceleration since he began tracking it.

Read the caveats before quoting the number, because they cut in an interesting direction. This is panel data, not a census of Google's logs, and it covers the United States only. It also excludes Google's mobile search app, where Fishkin expects the zero-click features fire harder still. So 68% is a conservative floor, not a precise constant.

Two independent teams, two unrelated methods, one direction of travel. That convergence is the durable finding; the specific decimal is not. Anyone quoting a figure to three places as settled fact is describing their citation habits, not search. What this looks like on a small service business is covered in zero-click search and the local business.

Change two: why did ranking stop predicting whether you get cited?

Because the engines stopped retrieving against your query and started retrieving against a spray of sub-queries they generate themselves. Ahrefs put numbers on the consequence in March 2026: across 863,000 keyword SERPs and 4 million AI Overview URLs, only 38% of AI Overview citations came from pages ranking in the top ten. Roughly 31% came from positions 11 to 100, and another 31% from pages beyond position 100 or not ranking at all. Eight months earlier that first figure had been near 76%.

The gap widens outside Google. In an earlier Ahrefs sample of 15,000 long-tail queries put to ChatGPT, Gemini, Copilot and Perplexity, just 12% of the URLs those assistants cited ranked in Google's top ten for the original prompt. Perplexity was the outlier at 28.6%, which fits its more conventional search-and-summarise architecture.

Ahrefs attributes the collapse to query fan-out, where one prompt expands into many hidden sub-queries, the best passage per sub-query is retrieved, then fused. That fits the data and matches how Google has publicly described AI Mode. It is still an inference drawn from outside the system. Neither Ahrefs nor we can see the retrieval logs, the generated sub-queries, or the weights picking one passage over another. Treat it as a well-supported working model, not a documented fact. More depth in AI search versus traditional SEO and how AI search differs from Google.

Change three: which query classes actually lost the clicks?

Informational queries absorbed nearly all of the damage, and this is the most useful fact on the page. Semrush's October 2025 sample of AI Overview keywords found 57.1% carried informational intent, against 18.57% commercial, 13.94% transactional and 10.33% navigational. AI Overviews appear where a question can be answered in a paragraph, precisely where an explainer article used to earn its click.

Seer Interactive measured the damage inside that class. Across 3,119 informational terms, 42 organisations and 25.1 million organic impressions between June 2024 and September 2025, organic click-through on AI Overview queries fell from 1.76% to 0.61%, a drop of 61%. The finding most commentary skips: on queries with no AI Overview, click-through still fell from 2.72% to 1.62%, a drop of 41%. Behaviour shifted whether or not the feature was on screen.

A caution on that table: Semrush's percentages describe which keywords trigger AI Overviews, not how much traffic each class lost, and Seer's figures come from 42 organisations weighted toward informational terms. Neither is a census. They are the best public segmentation available, which is a different claim.

What does Google itself say about all of this?

Google's official AI features documentation says there is no separate discipline to learn. Its position: optimizing for generative AI search is optimizing for the search experience, and is therefore still SEO, because the generative features run on the same core ranking and quality systems. That is first-party guidance from the company operating the surface, which makes it the strongest single citation here.

If you want the terminology sorted before briefing anyone, we untangle the acronyms in GEO versus SEO. The short version: the split describes where you measure, not two separate crafts.

What does nobody outside these systems actually know?

The retrieval and selection internals, which is the part everyone most wants explained. Overclaiming is the standard failure in writing about AI search, so the limits are worth naming precisely.

Take the top-ten overlap question from earlier. Ahrefs measured 38%. BrightEdge, running its own methodology, has published figures well outside that range. These studies are not lying to each other. They sample different keyword sets, define a citation differently, measure in different months and treat SERP features inconsistently. When several outside labs point different instruments at a probabilistic black box, a spread of answers is expected. The honest summary: top-ten overlap has fallen substantially since mid-2025, and the precise figure depends on who is counting.

Three further limits. Ahrefs' own authors note AI Overviews are probabilistic and shift between identical queries, so any snapshot is one draw from a distribution. No engine publishes retrieval logs, citation-selection weights or its generated sub-queries, so every published mechanism is reverse-engineered from outputs. And the panel studies behind the zero-click headline exclude surfaces where behaviour likely runs stronger: direction is trustworthy, magnitude soft.

What follows is a posture, not a shrug. Anyone selling a fixed formula for AI citation is modelling a system they cannot observe. What you can observe is your own Search Console data, your branded search volume, and what four assistants say about your business today. That is where the diagnosis lives, and what the self-check below uses.

So what actually did die?

Specific practices died, and they were the ones already running on thin ice. Several were genuinely load-bearing, so name them precisely.

• The shallow explainer ranking on a question answerable in two sentences. The click that funded that page has largely gone.
• Publishing to occupy a keyword rather than to say something. When an engine can synthesise a better answer from three stronger pages, a fourth restatement earns nothing.
• Reporting rank position as though it were a business outcome. Position and traffic measurably decoupled. Improved rankings alongside falling sessions is no longer a contradiction; it is the normal condition.
• Treating session volume as the only scoreboard. If unclicked exposure now does part of the introduction work, counting only visits understates the content.

What survived got more valuable, and Seer's dataset puts a figure on it: brands cited inside an AI Overview saw roughly 35% higher organic click-through than uncited brands on the same queries. Presence in the answer is not a consolation prize. It now decides whether the remaining clicks come to you.

Five questions to run against your own numbers

Run these in order and within an afternoon you will know whether AI search hurt you, and which part of your content took it. All of it uses tools you already have.

A pattern we hit repeatedly in audits: a business arrives certain its rankings collapsed, and the chart shows rankings held steady while only the click line bent. Different problem, different remedy, and the two get confused constantly because most dashboards put position and sessions on separate screens.

Does this land differently on an Ontario local business?

Yes, and mostly in your favour. Navigational and transactional queries are the two least exposed classes in the Semrush breakdown, together under a quarter of AI Overview keywords, and those carry local commercial intent. Someone typing "emergency plumber Brantford" at nine on a Sunday night gets a map pack and a phone number, not an essay. The generated answer has nothing to add and Google largely knows it.

Where a local operator feels it is one layer up the funnel. The educational content around the service is exposed: what a repair typically costs, how to tell whether you need one, what the warning signs are. Those pages used to bring in the researching homeowner three weeks before the call. Many now answer in the results and send nobody. If your lead flow leaned on that content to do the introducing, the loss is real even though your money queries look untouched.

What actually earns a citation, according to the research?

The clearest peer-reviewed answer comes from the paper that named the field. Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande introduced Generative Engine Optimization at KDD 2024, testing content changes against generative engines on a purpose-built benchmark and reporting visibility improvements up to 40% for the strongest methods. The tactics carrying that lift were editorial, not technical: adding quotations, adding statistics, citing authoritative sources. Keyword-density work did not help.

Two caveats keep that honest. The paper's own conclusion is that effectiveness varies by domain, so 40% is a ceiling from favourable conditions, not a number to promise anyone. And it tested engines as they existed in 2023, several model generations back. Direction has held up better than magnitude, the normal fate of empirical results in a fast-moving field.

What matters is that two independent sources point the same way. The academic work says quotations, statistics and citations improve visibility in generated answers. Google's documentation asks for original information, first-hand expertise and a real point of view. Neither describes a trick. Both describe writing something a summary cannot replace.

The verdict, stated once more

No, SEO is not dead, and the question has cost more attention than the answer is worth. Three things changed, all measurable: roughly two in three United States Google searches now end without a click, ranking predicts citation far less reliably than a year ago, and the losses land almost entirely on informational queries while branded and local demand holds. If your pipeline ran on explainer content, you have real work ahead. If it runs on branded and local search, you are better insulated than the headlines suggest.