AI Overviews and GEO, minus the snake oil.

The names, quickly

AIO means AI Overviews, the machine-written summaries Google puts above its results. GEO means Generative Engine Optimization, the work of getting your business cited inside AI answers, whether that answer comes from Google, ChatGPT, Perplexity, Claude or Copilot. You’ll also hear AEO (answer engine optimisation), LLMO, AISO and a few others. They all describe the same job. The industry has spent two years arguing about the label and nobody has won, so we say GEO, half our clients say AEO, and Google would quietly prefer nobody said either. More on that later.

What AI Overviews are, and how big they’ve become

An AI Overview is a summary generated by Google’s Gemini models and stitched together from pages it retrieves at the moment you search. It sits above everything else, cites a handful of sources in cards to the side, and answers the question well enough that most people stop there.

The rollout has been quick. Google trialled the idea as SGE in 2023, launched AI Overviews in the US in May 2024, reached the UK that August, and by 2025 they were live in over 200 countries. In January 2026 Google made Gemini 3 the default model behind them, and within weeks around 42% of the domains being cited had changed, which tells you a lot about how settled any of this is.

How often you’ll meet one depends entirely on what you search. Google’s own line in early 2026 was that AI Overviews appear on roughly half of US queries. Independent trackers put it anywhere from a fifth to a half depending on the keywords they sample, and UK panels tend to sit lower, around the 20% mark. The pattern underneath is more useful than the headline: health and other heavily informational topics trigger them most of the time, long conversational questions are far more likely to get one than short keywords, and straightforward shopping queries mostly still don’t. The mix is shifting, though. Commercial and even brand searches now trigger them far more than they did in early 2025.

Sitting alongside all this is AI Mode, a separate tab that turns Google into a full conversation. It reached the UK in July 2025 and works differently under the bonnet: your question gets split into a batch of related searches (Google calls this query fan-out), the results are read, and one answer comes back with links. Google has been nudging people from Overviews into AI Mode through 2026, and ads are already appearing inside both. Direct traffic from it is hard to isolate, which we’ll get to.

The click problem

Here is the part that hurts. Pew Research watched about 69,000 real American searches and found that when an AI summary appeared, only 8% of visits ended with a click on a result. Without one, 15% did. Roughly one visit in a hundred clicked a source cited inside the summary itself.

Other studies point the same direction with different numbers. Ahrefs measured a 34.5% fall in click-through for the top organic result when an Overview shows. Seer Interactive, tracking 2.4 billion impressions across 53 brands, watched organic click-through on Overview-carrying queries collapse from 1.76% to 0.61% during 2025, then recover to about 2.4% in early 2026 after Google shortened the summaries and pushed exploration into AI Mode. Even after that recovery, queries with an Overview still convert clicks at roughly a third below queries without one. Across everything, around two thirds of Google searches now end without a click to any website.

Two things stop this being a funeral. First, impressions have gone up sharply even as clicks fell, because your brand is now being shown (and sometimes quoted) to people who never visit. Second, the clicks that survive are better. Adobe’s retail data had AI-referred visitors converting 38% worse than average in March 2025 and 42% better a year later. Shopify reports AI-referred sessions converting around 50% higher than organic search on product pages. The pattern varies wildly by sector, and some B2B categories see no lift at all, but the direction is clear enough: fewer clicks, warmer clicks. If you measure sessions, this era looks like decline. If you measure revenue, it often doesn’t. You can probably guess which side of that argument we’re on.

Where GEO came from, and what Google thinks of it

The term has an unusually specific birthday. In November 2023, researchers from Princeton, Georgia Tech and the Allen Institute published a paper called GEO: Generative Engine Optimization, testing nine ways of rewriting content to see what earned more visibility in AI-generated answers. The winners were adding quotations, statistics and cited sources, worth roughly 30 to 40% more visibility in their benchmark. Keyword stuffing did nothing. It was slightly worse than doing nothing, in fact, which felt like the universe finally being fair about something.

Take the exact percentages with some salt, because the paper tested Perplexity-style setups rather than Google, and real engines have moved on since. But it remains the best public evidence that how you write changes whether machines quote you, and its core finding (specifics get cited, padding doesn’t) has held up in practice.

We’d call that true and incomplete. True, because the fundamentals genuinely do dominate, and an ocean of GEO snake oil deserved the bucket of cold water. Incomplete, because Google’s account covers Google, and ChatGPT, Perplexity and Claude play by their own rules on crawling, freshness and sources. Incomplete even about Google’s own feature, honestly, because independent tracking finds only around 17% of AI Overview citations now come from pages ranking in the organic top ten, down from roughly three quarters in mid-2024. Ranking still matters enormously. It just stopped being the whole story.

How the machines choose what to say

Every AI answer is built from two ingredients, and they reward different work.

The first ingredient is memory: whatever the model absorbed about the world, and about you, during training. This updates on a lag of months. It is why a model can describe a well-known brand fluently with no search involved, and why a business nobody writes about is invisible to it. You influence this route by being mentioned, described and reviewed across the web, in text the training crawlers are allowed to read. Links are optional here. A model doesn’t need a followed link to learn that a Manchester agency exists and what it’s known for. It needs the sentence to exist somewhere it read.

The second ingredient is live retrieval. When the question needs current facts, the engine runs real searches and reads the results before answering. AI Overviews and AI Mode retrieve from Google’s index. ChatGPT’s search leans on Bing’s index plus OpenAI’s own crawler. Perplexity maintains its own. This route rewards pages that rank, load, and hand over a clean quotable answer, and it can surface content published yesterday.

Almost everything that follows is one of those two jobs: be known, so the memory route works for you, and be fetchable, so the retrieval route can find and quote you. And given that a single model swap reshuffled 42% of Google’s citations in a few weeks, treat any tactic aimed at one engine’s current habits as rented ground. The two jobs are the freehold.

What actually works

1. Keep ranking. Boring, but it’s the engine room

Retrieval starts from search indexes, so a page that can’t rank can’t easily be found to quote. The overlap between citations and rankings has loosened, as above, but loosened is a long way from gone: a strong organic position remains the single most reliable way into an AI Overview, and into anything that retrieves through Bing or Google. Everything on our SEO services page still applies. Sorry. We checked.

2. Decide which bots you let in

This is now a real commercial decision, made in robots.txt, and the bots are separate levers. Googlebot feeds Search and therefore AI Overviews and AI Mode; block it and you vanish from Google entirely, so that one’s not really a choice. Google-Extended only controls whether your content trains Gemini models, and blocking it does not remove you from AI Overviews, which surprises nearly everyone. GPTBot is OpenAI’s training crawler; OAI-SearchBot is what lets ChatGPT’s live search retrieve you; ClaudeBot and PerplexityBot follow the same logic for Anthropic and Perplexity. Blocking training bots protects your content a little and costs you presence in the models’ memory. There’s no right answer, but there is a wrong one, which is not knowing what your file currently says. Our robots.txt generator builds the rules either way, and if you want a page kept out of AI summaries while staying indexed, the nosnippet and max-snippet controls do exactly that.

3. Put the answer first, then earn the rest of the page

Retrieval systems pull passages, and a passage travels best when it makes sense on its own: the question as a heading, the answer in the first sentence or two, a number in it, caveats afterwards. Google’s guide says you don’t need to restructure content for machines, and strictly it’s right, because this was already how good featured-snippet writing worked years before anyone said GEO. Write the version a stranger could lift out whole and you’ve written the version both humans and models prefer.

4. Give them numbers worth stealing

The original GEO paper’s clearest finding was that statistics, quotations and cited sources earn visibility. Models reach for specifics because specifics sound like answers. The strongest version of this is data nobody else has: your own benchmarks, your own survey, a properly maintained resource. Our algorithm update timeline exists partly for this reason; reference material gets cited long after opinion pieces are forgotten. If you publish one number the whole industry repeats, the machines will repeat it too, with your name attached.

5. Make it unmistakably clear who you are

Models deal in entities, and confusion about yours is expensive: same name, address and phone everywhere, a proper about page, organisation schema with sameAs links tying your site to your profiles. Structured data helps machines parse and disambiguate all of this, and it still wins rich results, so do it properly (our schema generator and schema guide cover the lot). Just hold the honest frame: Google’s guidance is explicit that no special markup buys AI visibility. Schema is plumbing. Plumbing matters. Nobody ever ranked because of a nice boiler.

6. Get mentioned where the models read

This is the memory route, and it looks suspiciously like digital PR because it is. Coverage in news and industry publications, presence in the roundups people actually consult, expert commentary with your name on it. The citation data makes the venues unusually concrete: by various counts YouTube appears in over a fifth of AI Overview citations, Reddit around a fifth, Wikipedia not far behind. Models trust places where humans argue and correct each other. Earn your way into those conversations. Astroturfing them is a different matter; Reddit’s users can smell a marketing account from another subreddit, and the whole point of the venue is that fakery gets found. One genuine, useful answer from a named founder outperforms fifty planted ones. This is the same authority-building we sell as digital PR, aimed at a new kind of reader that never sleeps.

7. Stay fresh where it counts

Retrieval engines carry a bias toward current material, and answer engines love a dated, recently-updated source. That means real maintenance of the pages that earn you money and citations: refreshed figures, honest revision dates, a changed year only when something actually changed. Faking the date while the page rots is the sort of thing that works right up until it’s the reason nobody trusts your dates.

8. Remember Bing exists

Strange sentence to type, but ChatGPT’s live search leans on Bing’s index, which quietly made Bing relevant again after a decade of being the punchline. Verify the site in Bing Webmaster Tools, submit the sitemap, consider IndexNow for instant pings. It’s an afternoon of work for access to the retrieval layer of the most-used assistant on earth. Cheap ticket.

Things we’d skip

llms.txt. A proposed file listing your key pages for AI systems. Lovely idea, politely ignored: Google has said on record it doesn’t use it, adoption sits around one site in ten, and when SE Ranking modelled 300,000 domains they found no relationship between having the file and getting cited. One monitoring firm watched half a billion AI bot visits and counted 408 requests for it. If you run developer documentation that coding agents consume, it has a genuine niche. For a normal business site it’s a lottery ticket that costs ten minutes; buy it if you enjoy lottery tickets, just don’t confuse it with strategy.

Hidden instructions and injection tricks. White-on-white text telling the AI your company is wonderful, prompts buried in markup. These occasionally work for a fortnight, then get patched, and the screenshots live forever. Being the case study in a “brands caught gaming AI” article is negative GEO.

Machine-written content at volume, aimed at machines. Models quoting models produces a grey slurry the filters are getting rapidly better at catching, and Google’s helpful content systems already gutted the thin-affiliate web once. If AI helps you draft, fine, so long as a human with actual expertise owns what ships. Publishing a thousand pages nobody checked is how sites end up on the wrong side of our algorithm timeline.

Anyone guaranteeing citations. Nobody controls what a model says. People selling certainty in this market are selling the certainty, not the result.

How to tell if any of it is working

Awkwardly, the biggest surface is the hardest to see. Google folds AI Overview and AI Mode clicks into ordinary Search Console totals, so a common 2026 signature is impressions climbing while clicks sag, which is your brand appearing in answers people don’t leave. A dedicated generative-search report has been rolling out in Search Console since June 2026, UK properties first, so check whether your account has it yet; first-party data beats guesswork.

The assistants are easier. Referrals from chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com show up in GA4 if you group them into a channel, with ChatGPT typically supplying the overwhelming majority. The percentages look tiny next to organic. Judge them on conversion rate and revenue per visit before deciding they’re tiny in the way that matters, and know that plenty of AI traffic arrives with no referrer at all and hides in Direct, so whatever you measure is an undercount.

And do the unglamorous version: write down the twenty questions that lead to money in your world, ask them monthly in ChatGPT, Perplexity, Gemini and an AI Overview, and log who gets named. Half an hour, a spreadsheet, no subscription. When a client asks what their GEO position is, that spreadsheet is a more honest answer than most dashboards, and watching a competitor’s name appear where yours should be is a wonderfully clarifying experience.

So how worried should you be?

Depends what you sell. If your business is pageviews on informational content (recipes, definitions, “what is” articles, affiliate roundups), the squeeze is real, well documented and unlikely to reverse; that traffic is being answered in place. If you sell services, products or local expertise, the picture is gentler: buying queries trigger fewer Overviews, people still click to hire and to purchase, and the visitors AI sends arrive unusually ready. The businesses coping worst are the ones who measured success in sessions. The ones coping best own their niche’s answers thoroughly enough that the machines keep reaching for them.

Which brings us to the slightly embarrassing conclusion: after all the acronyms, the best generative engine optimisation looks like ranking well, writing like you mean to be quoted, earning coverage in places that matter, and measuring money instead of traffic. Either that’s reassuring or it’s annoying, depending on how you’ve spent the last decade. We find it reassuring, but then we would.

Questions we keep getting asked