AI SEO

The two meanings of “AI SEO”

Search for “AI SEO” and you will land on two entirely different kinds of page without anyone flagging the switch. One set is about software that writes your meta descriptions. The other is about not disappearing from the internet. Both are legitimate; they solve different problems and cost different amounts of money.

The second meaning is winning the term, and the scale is why. Google reported in June 2026 that AI Overviews has over 2.5 billion monthly active users and AI Mode has passed one billion monthly users; OpenAI reported more than 900 million weekly active users in February 2026. Similarweb's panel puts generative AI platforms at 9.5 billion monthly visits, up 70% year on year, with ChatGPT's share of that traffic falling from roughly 76% to 53% between June 2025 and May 2026 as Gemini climbed past 27%. That last number matters more than the headline: the surface is fragmenting, so “are we visible in AI?” is already the wrong question. The question is which engine, and the answer differs by engine. Even Wikipedia has now settled the vocabulary in the same direction — its “AI SEO” entry redirects to Generative Engine Optimization.

Now the number that almost every page in this category leaves out, and the one we would rather you knew before you spend anything. Semrush measured the traffic mix across more than 50,000 websites and found that AI traffic grew 66% during 2025 but still amounted to just 0.14% of all visits, while organic search held 16.04% of traffic and grew 2.38%. That is published by a company selling AI-visibility tooling, which is what makes it worth trusting. Read it carefully, because it does not say what either camp wants: AI is not yet a meaningful traffic channel, and it is already a meaningful decision channel. With AI Overviews at 2.5 billion monthly users and Pew measuring clicks halving when a summary appears, the buyer forms a shortlist inside the answer and arrives later through a channel your analytics will label direct or branded search. If you justify AI SEO on referral sessions, the numbers will not support you and you will be right to stop. The case for it is influence at the point of decision, and that has to be measured where it happens.

What sits under AI SEO: the terminology, mapped

Four names circulate for overlapping work, which makes the category look more confusing than it is. The honest hierarchy is that AI SEO is the umbrella and the rest are specialisms inside it:

• GEO — Generative Engine Optimization. Being mentioned, cited and recommended inside answers that generative engines write. The brand-level layer. See what is GEO.
• AEO — Answer Engine Optimization. Writing self-contained passages an engine can lift verbatim as the answer. The page-craft layer. See answer engine optimization.
• LLM SEO. The technical plumbing: server-rendered HTML, crawler access, entity clarity, retrievability. The layer everything else stands on. See LLM SEO.
• AI search optimization. The execution playbook across every AI search surface at once. See AI search optimization.
• Classic SEO. Still underneath all of it, and still feeding it — see GEO vs SEO for the full side-by-side.

Nobody will be fired for saying AEO when they meant GEO. What matters operationally is the level you are working at: AEO is something you do to a page, GEO is something you do to a brand, LLM SEO is something you do to a codebase, and AI SEO is the budget line all three come out of.

Does the category even exist? What Google says

It is worth confronting the strongest argument against this entire field, because almost no page ranking for “AI SEO” does. Google publishes an official guide to AI-features optimization, last updated 10 July 2026, and its position is that there is nothing new here: “optimizing for generative AI search is optimizing for the search experience, and thus still SEO”. The same guide tells site owners explicitly not to create llms.txt files or special AI markup — “Google Search ignores them” — not to chunk content into fragments for machines, not to rewrite pages specifically for AI systems, not to pursue inauthentic mentions, and not to over-focus on structured data, since “structured data isn't required for generative AI search.” Its companion page is blunter still: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.”

Take that seriously — it is free, first-party, and it invalidates a meaningful slice of what the GEO advice industry sells. Independent data agrees on at least one point: Ahrefs tracked 1,885 pages that added schema markup and found AI citations barely moved, since most models never see your JSON-LD — it is stripped before the text reaches them.

Two things Google's framing does not cover, though. First, scope: Google is describing its own AI Overviews and AI Mode. On those surfaces, 37.9% of cited URLs also rank in Google's top 10, so “it's still SEO” is a fair description. Across ChatGPT, Gemini and Copilot the overlap with Google's top 10 falls to roughly 12%, and Google has no visibility into, or advice about, those engines at all. Second, measurement: even if you accept that the work is identical, the reporting is not. Nothing in Google's stack tells you whether ChatGPT recommended you this morning. That gap is the honest reason this category exists.

The AI SEO workflow, end to end

AI SEO is not a separate channel bolted onto the side of your marketing. It is one loop with six stages, each of which changes in a specific way once answers, not links, are the output.

1. Demand research: keywords become prompts

Classic keyword research targets the phrase a person types. Generative engines rewrite that phrase into several related sub-queries before they search anything, so you are no longer competing for one string — you are competing for a neighbourhood. The inputs have physically changed shape: Google reports that the average AI Mode search is triple the length of a traditional Search query, and that AI Mode queries have more than doubled every quarter since launch. Practically, this means building a prompt portfolio rather than a keyword list: the natural-language questions your buyers would actually ask an assistant, tagged by funnel stage, with the real search volume behind each so you know which ones are worth winning. Our study of GEO search demand works through what that demand actually looks like.

2. Access: can the crawlers reach you at all

This is a hard gate with no partial credit. Check robots.txt and your CDN bot rules for GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot and Google-Extended. A blocked crawler is not a small penalty; it is a guaranteed zero on that engine. Worth knowing the economics before you decide: Cloudflare's analysis found Anthropic crawling roughly 38,000 pages per referred visitor in July 2025, OpenAI about 1,091 and Perplexity 195, against Google's 5.4. Cloudflare flags its own caveat, which most people repeating this stat drop: some assistants send no referrer header, so the real gap is narrower by an unknown amount. Even discounted, the direction is unambiguous — AI engines read far more than they refer. Blocking training-only crawlers is a defensible business decision; blocking the search and user-agent crawlers is self-harm, and the two are different bots.

3. Rendering: the failure mode with no SEO equivalent

Vercel and MERJ analysed more than 500 million AI-crawler fetches in December 2024 and found that none of the major AI crawlers execute JavaScript — GPTBot requests JavaScript files in about 11.5% of its fetches but never runs them. Googlebot renders your client-side app; OpenAI's and Anthropic's crawlers see the shell. A site can rank perfectly on Google and be functionally blank to an answer engine. Server-side rendering or static generation for your key pages is the highest-leverage technical fix in AI SEO. More on this in LLM SEO.

4. Content: answer-first, and evidence-dense

Engines retrieve passages, not pages, so every section has to survive being lifted out of context. Put the direct answer in the first one or two sentences at every level. Then add evidence: the Princeton-led GEO study, benchmarked over 10,000 queries, found that adding statistics lifted a source's visibility in generated answers by around 41% and citing sources lifted it by up to 115% for mid-ranked pages, while keyword stuffing performed roughly 10% worse than an unoptimised baseline on Perplexity (Aggarwal et al., GEO, KDD 2024). That study runs in a controlled five-source harness, so treat the magnitudes as directional rather than as a promised return — but the ordering of the levers has held up.

One thing to stop doing: chasing rich results that no longer exist. Google stopped showing FAQ rich results on 7 May 2026 — announced only in the Search Central documentation changelog, never in a blog post — and removed the documentation the following month. HowTo rich results went earlier: restricted to desktop on 8 August 2023, then deprecated outright on 13 September 2023. Writing a clean Q&A block is still good practice — each answer is a naturally self-contained, extractable unit — but do it for the reader and the retrieval, not for a snippet that was switched off.

5. Off-site presence: the biggest lever, and the least like SEO

Ahrefs' December 2025 study of 75,000 brands found branded web mentions were the strongest measured correlate of AI visibility on every surface — 0.709 for Google AI Mode, 0.664 for ChatGPT, 0.656 for AI Overviews — against roughly 0.194 for backlink counts, with YouTube mentions highest of all at about 0.737. These are correlations, not proven causation, and mentions travel with quality. But the ordering has now held across two runs of the study, and it points somewhere specific: the roundups, directories, review sites and community threads the engines already cite for your category are worth more attention than another link build. Being accurately described on pages you do not own is the work.

6. Measurement: three instruments, none of them GA4

You cannot manage this from one dashboard, and anyone selling you a single number is selling you a simplification:

• Search Console's Generative AI performance report — Google's own recommendation for measuring how your content performs in its generative AI features. Free, first-party, and limited to Google's surfaces.
• An AI visibility tracker for everything Google does not run. This works by re-asking your buyers' prompts on a schedule and counting mentions, so it is a sample — which is why every number needs a confidence interval and why blending engines into one score hides the case that matters.
• Server or CDN logs for crawler behaviour. AI crawlers do not execute JavaScript, so GPTBot, PerplexityBot and ClaudeBot never fire a GA4 tag. If you are looking for them in GA4, you will conclude they never came.

And set the right KPI. Pew Research, tracking 900 US adults across 68,879 searches in March 2025, found users clicked a result 8% of the time when an AI summary appeared, against 15% when one did not, and Ahrefs' February 2026 analysis of 300,000 keywords found AI Overview presence correlating with a 58% lower average click-through rate. If your AI SEO KPI is sessions, you will conclude that AI SEO does not work. The outcome you are buying is being named in the answer.

Hold that evidence honestly, though, because it is contested. Seer Interactive analysed 5.47 million queries and 2.43 billion impressions across 53 brands and found AI Overview organic CTR recovering from 1.31% in December 2025 to 2.36% in February 2026, with cited brands earning about 120% more clicks per impression than uncited ones — though still around 38% below a no-AIO baseline. Google, for its part, says total organic click volume has been “relatively stable year-over-year”, though it published no sample or method alongside that claim, so treat it as a position rather than a measurement. The synthesis that survives all three: the click is smaller, being in the answer recovers much of it, and nobody outside Google knows the true magnitude.

Using AI to do the SEO work

The other half of the term. AI genuinely helps with clustering and intent classification at scale, first-draft briefs and outlines, internal-link suggestions, structured-data generation, log-file and crawl analysis, and translating a keyword list into natural-language prompts. What it does not do is produce publishable authority.

The evidence here is unusually clear. Ahrefs analysed 1,000,000 pages from the top 10 across 100,000 SERPs in June 2026 and found that 9% of top-ranking pages were at least 80% AI-written and 5.3% were detected as fully AI-generated — so AI content is plainly not disqualified. But the same study found low- and moderate-AI pages receiving two to three times the impressions of high-AI pages, and indexation running 49.28% for low-AI pages against 40.35% for very-high-AI ones. Google is not punishing AI; it is punishing the thin, undifferentiated output that tends to accompany it. Google's own guidance is consistent with that, warning against generating content variations at scale primarily to manipulate rankings.

The working rule: use AI for the parts of SEO that are structured and verifiable, and keep a human on the parts that require judgement, original data or first-hand experience. Those are also, not coincidentally, the parts that make a page worth citing.

What changes, and what doesn't

Unchanged: crawlability, clean site structure, page speed, factual accuracy, genuine topical depth, and being a source other people reference. All of it pays out on both surfaces, which is why Google can reasonably say the work is still SEO.

Changed: the unit of competition (a passage, not a page), the unit of success (a mention, not a click), the strongest off-site signal (mentions over links), the technical floor (no JavaScript rendering), the measurement model (a sample, not a census), and the shelf life of a win (days to weeks, and reversible).

Retired: keyword density, rank-one tunnel vision, FAQ and HowTo rich-result chasing, and the assumption that your analytics can see your acquisition channel.

A 90-day AI SEO plan

1. Weeks 1–2 — baseline everything. Search Console for classic and Google-AI performance, an AI-visibility check for the other engines, and a server-log pull for AI crawler hits. You cannot manage a mix you have not measured.
2. Weeks 3–4 — clear the technical gate. Server-render your key pages, fix the robots.txt and CDN bot rules, clean up 404s and redirect chains. This is the work with the shortest path to effect.
3. Weeks 5–8 — rebuild 20 pages answer-first. Direct answer in the opening two sentences of every page and section, one real statistic with a named source and link per major claim, no keyword padding.
4. Weeks 9–12 — go and get mentioned. Identify the third-party pages the engines already cite in your category and get accurately represented on them. This is the slowest lever and the one with the strongest correlation behind it.
5. Ongoing — run the loop weekly. Re-sample the same prompts, compare against the confidence interval rather than the point estimate, and refresh your top pages on a real cadence.

For the engine-specific version, see how to rank on ChatGPT. For the tools at each stage, see AI SEO tools. For the tactical execution sequence, AI search optimization.

Where steek fits

steek covers the measurement stage for the engines Google does not run. It tracks how often ChatGPT, Perplexity, Gemini and Google AI Overviews mention, cite and recommend you, reports a 95% confidence interval on every metric so a small move is not mistaken for a result, shows the real monthly search volume behind each tracked prompt, and runs the AI-crawler and robots.txt checks described in stage two. For online stores it tracks per-SKU AI Shopping visibility and ties it to real revenue in GA4. It also shows the exact third-party sources your competitors are cited in that you are absent from — which is the shortlist for stage five. Plans start from €19/mo with every engine included, and the checker on the homepage runs free without a signup. For the wider landscape, see the best AI visibility tools.