Two kinds of reader land on this page. One has watched organic sessions fall for three quarters and needs to know whether AI search caused it. The other has been asked by a board, an investor or an agency what the company’s AEO strategy is, and does not want to answer badly.
Both questions have honest answers. Neither answer is the one being sold.
Last reviewed September 2026. Every figure below is dated in the sentence that carries it, because most of them will move.
SEO, AEO and GEO are not three disciplines
Start with definitions clean enough to repeat in a meeting.
SEO, search engine optimisation, is the practice of getting pages crawled, indexed and ranked in a list of links so a person clicks through. The unit of success is a position. The payoff is a session.
AEO, answer engine optimisation, is the practice of making content usable as a direct answer, whether that answer is a featured snippet, a voice assistant reply or an AI summary. The unit of success is inclusion in the answer.
GEO, generative engine optimisation, is the practice of making content more likely to be quoted or cited inside a generated response from a system such as ChatGPT, Google AI Mode or Perplexity. The unit of success is a citation.
SEO competes for a position in a list of links. AEO and GEO compete for inclusion in a block of text that something else generated. The unit of success changes from a rank to a citation.
| Term | Optimises for | Where it shows | How you measure it |
|---|---|---|---|
| SEO | Rank in a list | Blue links, product results, local pack | Impressions, position, clicks, sessions, revenue |
| AEO | Being the answer | Featured snippets, voice replies, AI summaries | Snippet ownership, AI impressions, referral sessions |
| GEO | Being cited inside a generated answer | ChatGPT, AI Mode, Perplexity, Copilot, Gemini | Prompt-set citation rate, sampled and averaged |
The inputs overlap heavily. The measurement does not overlap at all. That is the whole distinction, and it is enough to hold a conversation with anyone selling you a third retainer.
AEO has no founding document
Worth saying because vendors imply otherwise. There is no paper, no named author and no traceable coinage for AEO. It grew out of the featured snippet and voice assistant era and spread as a label after ChatGPT, and today most vendors use it interchangeably with GEO. It is a label, not a body of knowledge. If someone pitches an AEO methodology, ask what is inside it that is not SEO.
GEO does, and it is older than the pitch deck
GEO was introduced in “GEO: Generative Engine Optimization” by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at ACM KDD 2024 and first posted as arXiv:2311.09735 in November 2023. The headline everyone quotes is up to 40% visibility improvement. The part nobody quotes: the main evaluation ran on a simulated generative engine, GPT-3.5-turbo reading the top five Google results, not a live product. Validation on Perplexity produced smaller gains of 22% to 37%.
The findings are more useful than the headline. Adding quotations improved visibility by about 41%, adding statistics by about a third, citing sources by just under 30%, and fluency and readability work by roughly 15% to 30%. Keyword stuffing produced no improvement and performed worse than the untouched baseline.
Two conclusions. The direction is credible and it has aged reasonably well. And anyone quoting “+40% visibility” in 2026 as current methodology is quoting a 2023 lab result about a model nobody uses.
What actually changed
Not Google’s share. Google was still around 91% of worldwide search engine referrals in July 2026, at 91.32% against Bing’s 4.46% (StatCounter via Statista). The click changed, not the market.
SparkToro, working from Similarweb clickstream data, measured 68.01% of US Google searches ending without a click between January and April 2026, up from the 60.45% it recorded for 2024. Their panel excludes the Google mobile app, so the true figure is probably higher. Rand Fishkin also notes AI Mode was only 0.34% of searches in that window, which means AI Mode is not what drove the rise.
The most quoted figures, Pew’s, are older than people present them. Pew studied 900 US adults and 68,879 Google searches in March 2025, published July 2025. Users clicked a traditional result on 8% of visits where an AI summary appeared, against 15% where none did. Clicks on links inside the summary: 1% of visits. Cite them as March 2025 data or do not cite them.
There is exactly one randomised experiment on this. Agarwal (Indian School of Business) and Sen (Carnegie Mellon Heinz College) ran a Chrome extension that randomly assigned 1,065 US desktop users to see or not see AI Overviews, two weeks each, between 7 January and 10 February 2026 (SSRN 6513059). Conditional on an AI Overview appearing, outbound organic clicks fell 39.8% and zero-click searches rose 34.5%. Sponsored clicks were unaffected. They found no improvement in user satisfaction.
That last point is a live dispute, not a closed one. Liz Reid, who runs Google Search, stated in an official post on 6 August 2025 that total organic click volume from Search to websites has been relatively stable year over year and that click quality has improved. The experiment does not find the quality improvement. Both positions are on the record, and anyone presenting either as settled is arguing rather than reporting.
One more caution. Prevalence numbers for AI Overviews vary wildly between trackers because the keyword sets, geographies, devices and detection methods differ. BrightEdge‘s commercial tracker recorded 48% by March 2026. Semrush, across more than 600,000 US desktop keywords from November 2025 to April 2026, found AI Overview appearance on commercial-intent results up 71% while transactional-intent results went down 5%. Never quote one prevalence percentage as the number.
Finally, separate publishers from brands. Digital Content Next members reported traffic losses mostly between 1% and 25% across eight weeks in May and June 2025, and Penske Media sued Google over AI Overviews in September 2025. That damage is real and it proves the mechanism exists. It is not your exposure profile. A publisher sells the click. You sell a product, and the person who reaches your site after reading an AI answer has already been pre-qualified.
Two things everybody did that the evidence does not support
llms.txt. Ahrefs analysed 137,000 domains using its own web analytics, a sample already skewed towards SEO-aware sites: 28% of them had published one, and 97% of those files received zero requests in May 2026. Of the requests that did occur, 96% came from bots, led by SEO audit tools at 21.7%, and only 19.5% came from named AI tools. Google’s Gary Illyes confirmed in July 2025 that Google does not support llms.txt and has no plans to. John Mueller compared it to the keywords meta tag.
Schema retrofits. Ahrefs ran a matched difference-in-differences on 1,885 pages that added JSON-LD between August 2025 and March 2026 against roughly 4,000 control pages, published 11 May 2026. AI Overview citations moved by -4.6%, AI Mode by +2.4% and ChatGPT by +2.2%, the last two indistinguishable from zero. Read their caveats: only pages already receiving 100 or more AI citations were studied, so this does not prove schema is useless on a page with no visibility. It does show that bolting schema onto pages that already rank is not the lever.
Google says the same thing in its own documentation, in the guide to optimising for generative AI features published in May 2026:
You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them.
The same guide states there is no special schema.org markup for generative AI search. It does name Merchant Center feeds and Google Business Profile as relevant, and it requires pages to be indexed and snippet-eligible. Ranking still helps, but less than it did. In July 2025 Ahrefs found 76.10% of pages cited in AI Overviews also ranked in the organic top 10, with the primary cited URL sitting at a median position of 2. In the rerun they published on 2 March 2026, across 863,000 keywords and 4 million cited URLs, that share had fallen to 38%, with 31.2% of citations coming from pages ranking between 11 and 100 and 31.0% from pages ranking beyond position 100. Quote the 2026 figure, and date it.
Sort the work by who controls the input
This is the reordering that makes the topic tractable, and it explains why almost nobody writes it. Two of the three buckets cannot be sold as a monthly retainer.
1. Your server. One afternoon, mostly free.
The distinction almost nobody gets right: GPTBot is OpenAI’s training crawler, OAI-SearchBot is the crawler that makes your pages eligible to appear in ChatGPT’s search results, and ChatGPT-User is the agent that visits a page when someone asks ChatGPT to look at it. GPTBot and OAI-SearchBot take separate robots.txt rules. ChatGPT-User is user-initiated, and OpenAI’s documentation states that robots.txt rules may not apply to it. Blocking “AI bots” on legal advice can remove a brand from the answer layer without anyone in marketing knowing.
Cloro looked at 1,058 prominent domains in July 2026: median citations per Google ranking were 0.003 for domains blocking GPTBot against 0.417 for domains allowing it, and zero for domains blocking OAI-SearchBot against 0.394 for those allowing it. The authors state plainly that this cannot prove causation. Even so, 13.9% of those domains blocked GPTBot and only 3.4% blocked OAI-SearchBot, which tells you most blocks were made without understanding the difference. Read your robots.txt today, and your CDN bot rules with it.
2. Your catalogue. This is where a consumer brand actually decides the outcome.
Shopify turned Agentic Storefronts on by default in the Winter ’26 Edition, announced 10 December 2025, naming ChatGPT, Perplexity and Microsoft Copilot as channels. The Spring ’26 Edition on 17 June 2026 added Shopify Catalog, a normalised structured product dataset, and the Universal Commerce Protocol developed with Google, both enabled by default for eligible merchants.
So the product data question is no longer whether you added FAQ schema. It is whether your variants, attributes, availability and prices are clean enough to survive being normalised into somebody else’s catalogue. We run these stores. Merchant feeds fail quietly and constantly, nothing on the site breaks when they do, and nobody notices for months.
The counterweight, because agentic commerce is not a straight line. OpenAI launched Instant Checkout in September 2025 on the Agentic Commerce Protocol built with Stripe, and retired it in March 2026 after roughly six months and around 30 live merchants according to Forrester, citing product data accuracy, inventory and multi-item carts. CNBC covered the pivot on 24 March 2026, back to a discovery-first model that routes shoppers to merchant storefronts. Which tells you where the durable investment sits: clean, findable product data, not checkout integrations.
3. Other people’s platforms. The uncomfortable one.
When AI engines recommend products, they mostly do not cite brand websites. Cloro tested 3,312 product-intent prompts across six engines, published 26 May 2026: the most cited sources were YouTube at 19%, Reddit at 19%, RTINGS at 16%, Google at 12% and Forbes at 9%. Brand-owned domains were, in the authors’ words, conspicuously absent. Amazon appeared in 5%. Peec AI‘s analysis of 30 million cited sources, reported in March 2026, found Reddit the most cited domain overall, then YouTube and LinkedIn.
Ahrefs’ correlation work across 75,000 brands points the same way. In their May 2025 study of AI Overviews, branded web mentions correlated at 0.664, Domain Rating at 0.326 and backlinks at 0.218. In their cross-platform study published in December 2025, covering ChatGPT, AI Mode and AI Overviews, YouTube mentions came out strongest at about 0.737, with branded web mentions between 0.66 and 0.71. Ahrefs state explicitly that this is correlation, not causation, and that every factor was moderate to very weak. So do not repeat “brand mentions are the new backlinks” as a mechanism. It is a pattern in a chart.
The pattern still has an implication you can act on. The budget line that moves AI product visibility looks like PR, creator seeding and getting properly reviewed by the two or three category reviewers your customers actually read. That is a distribution problem wearing an SEO costume, and it takes two quarters, not two weeks.
Size the opportunity revenue-weighted, not session-weighted
AI referrals are small. Conductor’s benchmark across 13,770 enterprise domains, using data collected between May and September 2025, put them at 1.08% of sessions.
They are not small in value. Adobe Analytics, working from more than a trillion visits to US retail sites, reported in August 2026 that in July AI-referred traffic converted 60% higher than non-AI traffic, the eleventh consecutive month of outperformance, with 53% more revenue per visit and 28% more cart additions. A year earlier the same traffic converted worse. The gap closed, then reversed.
The honest counterweight comes from Ahrefs, who popularised the conversion story. On their own site in June 2025, AI search was 0.5% of traffic and 12.1% of signups. Their study of 81,947 sites that same month found AI visitors weaker on engagement: 4.0 pages per visit against 5.2 from search, and a 67.8% bounce rate against 63.7%, though they stayed slightly longer, 86 seconds against 78. Both things are true. High intent, low exploration, one and done.
So the question for a brand doing a few million a year is not how much AI traffic you have. It is what a channel worth roughly 1% of sessions, converting well above average, justifies spending on. Usually something. Never everything.
What is measurable and what is still guesswork
Search Console now reports Search generative AI performance, announced 3 June 2026 with worldwide rollout completed 31 August 2026, covering AI Overviews and AI Mode. It gives impressions and only impressions. No clicks, no CTR, no average position, no query breakdown. Chart data aggregates by property, so page-level impressions will not sum to the property total.
GA4 added an AI Assistant default channel group in May 2026, applied automatically, assigning recognised AI referrers the medium ai-assistant. It is not retroactive and Google has not published the full referrer list. The blind spot matters: traffic arriving with no referrer header, which includes in-app browsers and pasted links, still lands in Direct. Your AI channel is undercounted and you cannot say by how much.
And every AI share of voice score is a sample of a non-deterministic system. LLM outputs vary between runs even at temperature zero, and as of September 2026 we know of no vendor-neutral cross-platform benchmark of the visibility tools. Track the direction of travel, act on it, and never present the second decimal place to anyone. That precision is invented.
What to fund this quarter
- Audit robots.txt and CDN bot rules for the GPTBot and OAI-SearchBot distinction. One afternoon, no vendor required, large downside if it is wrong.
- Fix product data at the source: variants, attributes, GTINs, availability, price accuracy, a Merchant Center feed with no silent failures. It pays in paid, organic and AI at the same time.
- Keep ranking. The overlap between AI Overview citations and the organic top 10 fell from 76% in Ahrefs’ July 2025 study to 38% in their March 2026 rerun, so ranking is no longer sufficient. It is still the cheapest way in: classic SEO is the entry ticket to the answer layer, not a competing budget.
- Fund earned coverage where the engines actually read: YouTube reviews, credible category reviewers, genuine community presence. Budget two quarters before you judge it.
- Own the channels that do not depend on a click at all. Email and SMS lists you own are the only hedge against every future ranking change.
- Set a baseline now: Search Console AI impressions, the GA4 AI Assistant channel, and a fixed set of 50 to 100 buying-intent prompts you re-run monthly.
What to refuse to pay for
- An llms.txt deliverable.
- A schema retrofit sold as an AI citation strategy on pages that already rank.
- Any pitch quoting the GEO paper’s 40% without mentioning it was a 2023 simulation on GPT-3.5.
- An AI share of voice dashboard reported to two decimal places with no confidence interval.
- Rewriting every H2 as a question because a listicle said so. Nobody has tested it.
How we work on this
EDEUS runs Shopify stores, product feeds, lifecycle flows and content in the same room, so we tend to treat this as one commercial problem rather than three acronyms. If you want a second opinion on what your current agency is charging you for, we are happy to look at it with you.
Frequently asked questions
What is the difference between SEO, AEO and GEO?
SEO optimises for a position in a list of links, and its payoff is a click. AEO optimises for being used as a direct answer, in a featured snippet, a voice reply or an AI summary. GEO optimises for being quoted or cited inside a generated response from a system like ChatGPT, Google AI Mode or Perplexity. The unit of success changes from a rank to a citation. The inputs overlap heavily; the measurement does not overlap at all.
Is GEO just SEO with a new name?
Not exactly, but it is closer than most vendors admit. GEO has a real origin: the paper “GEO: Generative Engine Optimization” by Aggarwal and colleagues, presented at ACM KDD 2024 (arXiv:2311.09735). AEO has no founding document at all and is an industry label from the featured snippet and voice assistant era. In practice the work that produces citations is mostly good SEO plus earned coverage on third-party platforms. Ahrefs found that 76.10% of pages cited in AI Overviews also ranked in the organic top 10 in July 2025; in their rerun published on 2 March 2026, across 863,000 keywords and 4 million cited URLs, that share had fallen to 38%. Ranking is a weaker entry signal than it was, which makes the earned coverage half of the job count for more.
Do AI Overviews actually reduce clicks?
Yes, on the only randomised evidence available. Agarwal (Indian School of Business) and Sen (Carnegie Mellon) randomly assigned 1,065 US desktop users to see or not see AI Overviews for two weeks each between 7 January and 10 February 2026 (SSRN 6513059). Conditional on an AI Overview appearing, outbound organic clicks fell 39.8% and zero-click searches rose 34.5%, with no measured improvement in user satisfaction. Google’s Liz Reid said in an official post on 6 August 2025 that total organic click volume has been relatively stable and click quality has improved. Everything else circulating on this topic is panel data or correlation.
Does llms.txt actually work?
There is no evidence that it does. Ahrefs analysed 137,000 domains using its own web analytics, a sample already skewed towards SEO-aware sites, and found that 28% of them had published an llms.txt file, and 97% of those files received zero requests in May 2026. Of the requests that did occur, 96% came from bots, led by SEO audit tools at 21.7%, with only 19.5% from named AI tools. Google’s Gary Illyes confirmed in July 2025 that Google does not support llms.txt and has no plans to, and Google’s own May 2026 guidance states you do not need to create AI text files or new markup to appear in Search.
Should I block GPTBot?
No, not by default, and the more expensive mistake is blocking the wrong one. GPTBot is OpenAI’s training crawler. OAI-SearchBot is the crawler that makes your pages eligible to appear in ChatGPT’s search results. ChatGPT-User is the agent that visits a page when someone asks ChatGPT to look at it. GPTBot and OAI-SearchBot take separate robots.txt rules, while ChatGPT-User is user-initiated and OpenAI’s documentation states that robots.txt rules may not apply to it. The same documentation says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, so that is the rule that removes you from those results. Cloro’s July 2026 analysis of 1,058 domains found median citations per Google ranking of 0.003 for domains blocking GPTBot against 0.417 for those allowing it, and zero for domains blocking OAI-SearchBot against 0.394 for those allowing it. The authors note this cannot prove causation. Blocking training while allowing search is a defensible position; blocking both by accident is the common mistake.
Does AI search traffic convert?
It is a small channel with unusually high value. Conductor’s benchmark across 13,770 enterprise domains, using data collected between May and September 2025, put AI referrals at 1.08% of sessions. Adobe Analytics, working from more than a trillion visits to US retail sites, reported in August 2026 that in July AI-referred traffic converted 60% higher than non-AI traffic, the eleventh consecutive month of outperformance, with 53% more revenue per visit. The counterweight: Ahrefs’ study of 81,947 sites in June 2025 found AI visitors viewed 4.0 pages per visit against 5.2 from search and bounced more often. High intent, low exploration.
How do AI assistants decide which products to recommend?
Mostly from third-party sources, not brand websites. Cloro tested 3,312 product-intent prompts across six AI engines, published 26 May 2026: the most cited sources were YouTube at 19%, Reddit at 19%, RTINGS at 16%, Google at 12% and Forbes at 9%, with brand-owned domains conspicuously absent from the top of the list. A separate Peec AI analysis of 30 million cited sources, reported in March 2026, found Reddit the most cited domain overall, followed by YouTube and LinkedIn. The practical consequence is that product visibility in AI answers is closer to a PR and creator seeding problem than an on-page optimisation problem.
Content created with the support of artificial intelligence
