AI Search & GEO
Sick of writing blogs? So are we. Here is what actually gets you recommended when your customer asks an AI instead of Google.
Most content marketing plans are a blog calendar with a strategy document stapled to the front. Twelve posts a quarter, chosen on keyword volume, published in the hope that traffic becomes enquiries. For a long time that worked well enough that nobody looked too hard at it.
It is breaking now, and not slowly. Your customer asks ChatGPT, Gemini or Copilot, or reads the AI answer sitting above Google’s results, and gets a usable answer without visiting anybody’s website.
The scale is easy to underestimate. In the first four months of 2026, 68% of Google searches ended without a click of any kind, up from 60% in 2024, and the fastest shift in a decade. Most of the demand you are competing for now resolves before anybody reaches a website.
So by the time someone does land on your site, they have already been handed a shortlist. The question is no longer whether you rank. It is whether you were on it.
You can’t rank your way into an AI answer any more
For years the honest advice was: rank well and everything else follows. That advice is now out of date, and the numbers are blunt about it.
In July 2025, 76% of the pages cited in Google’s AI Overviews also ranked in the top ten for that query. By March 2026 it was 38%. The remainder came from positions 11 to 100 (31.2%) and from beyond position 100 (31.0%). Pages that, by any traditional measure, were not visible at all.
Step outside Google and the link is weaker again. Across ChatGPT, Gemini, Copilot and Perplexity, only 12% of the URLs cited in AI answers rank in Google’s top ten for the original prompt. For ChatGPT, Gemini and Copilot on their own it sits between 6% and 8%. Four in five AI citations come from pages that do not rank anywhere in Google for the question being asked.
Rankings still matter. They are simply no longer the ticket.
So what decides it?
Fan-out. Before an AI answers you, it quietly breaks your question into dozens of smaller ones: background searches you never see. A question like “best wholesale nursery in Victoria” becomes queries about delivery radius, minimum order quantities, seasonal availability, trade accounts, substitution policy and a dozen more. The AI then assembles its answer from whatever it found across all of them.
Which means pages get cited because they turn up again and again across the sub-questions, not because they won the original one. Ahrefs attributes the decline above to exactly this behaviour.
Breadth across a topic now beats depth on a single term. That is the entire argument for building content hubs instead of blog calendars.
Being cited is not the same as being recommended
This is the trap most brands are currently walking into, and it is worth understanding before you spend a dollar.
Lily Ray at Amsive tracked 100 B2B “best [category] software” queries through Google’s AI Overviews across three months of 2026. Brands’ own self-promotional listicles, the “top ten tools in our category, ours first” page, were cited 323 times. But in 224 of those cases, roughly 69%, Google used the brand’s own page as a source and then recommended somebody else.
Read that again, because it is the whole problem in one sentence. Your page was good enough to inform the answer and not good enough to be the answer. The brands that did get recommended were the ones already widely referenced by third parties, not the ones with the best-optimised listicle.
You cannot self-promote your way into a recommendation. You get there by being genuinely useful across the whole question, and by being corroborated somewhere other than your own website.
Six things that actually move AI visibility
1. Breadth beats rank
Winning one keyword used to be the job. The job now is being present across the whole question. A page that appears in eight of the twenty sub-questions behind a query will beat a page that ranks first for the query and nothing else.
2. Follow the fan-out, not the keyword
Keyword research tells you what people type. Fan-out analysis tells you what the machine asks on their behalf before it answers. Those are two different lists, and the second one is where citations are won.
3. Don’t chase a fixed list of AI queries
Fan-out queries are probabilistic. Run the same search ten times and you will get ten different sets; run it on a different model and the overlap shrinks again. What you optimise for is the pattern that recurs, not the snapshot. Anyone selling you a stable list of your fan-out queries is selling you a screenshot.
4. Start where you are already losing
Open Search Console and find the pages with high impressions and falling clicks. That is AI answering the question for you, using your page, without sending you the visit. It is the cheapest priority list you will ever build, and you already own the data.
The losses are not spread evenly, which is what makes this worth doing properly. Amsive analysed 700,000 keywords and found that when an AI Overview appears, click-through rate falls 15.5% on average. It falls 20% for non-branded terms, and 27% for anything ranking outside the top three. Branded searches actually went up. The pain is concentrated in exactly the informational, non-branded territory most content programmes are built to win.
5. Put the question in the heading and the answer underneath
Use the question, or a close variant, as your subheading. Answer it in the very next sentence, not three paragraphs later. It is the most portable rule on this list and it costs nothing to apply to content you have already published.
6. Publish something only you know
Original data, your own numbers, what you have actually seen work. Everything else the AI can already get from a hundred other sources, and it will. Proprietary information is the only durable advantage in a system built to summarise everybody.
The PROOF framework
Once you know which questions matter, you have to decide what to build for each one. Most plans answer that with “an article”. That is the mistake.
PROOF maps the five jobs an AI-influenced buyer needs done, and the kind of asset that does each one.
| Stage | What is happening | What you need |
|---|---|---|
| P for Prompt | Someone asks the question, and the AI goes looking for source material. | Answer-led explainers, glossaries, trend pages, stats pages, short expert video |
| R for Recommend | The AI compresses what it found into a shortlist of three to five names. | Use case pages, comparisons, buyer guides, industry pages, alternative-to pages |
| O for Objection | The buyer checks whether the recommendation is actually true. | Case studies, methodology pages, expert bios, trust pages, FAQ hubs |
| O for Offer | They arrive on your site most of the way decided. | Pricing, scope, implementation, migration, calculators, commercial FAQs |
| F for Flywheel | Their experience becomes the next buyer’s evidence. | Reviews, customer interviews, original research, changelogs, partner pages |
The unit of work is a PROOF Pack: five assets for every priority topic, one per stage. Not five blog posts. Five different kinds of asset, because the five jobs are genuinely different.
Do that across your six most commercially important topics and you have a content programme built for how buying actually happens now. Do it as a blog calendar and you have thirty articles competing with each other for a click that increasingly never comes.
How we find the work
Deciding what to build first is a research problem, not a brainstorm.
- Find the topics you already have a foothold in. Search Console shows where you earn impressions but are losing the click. Those pages are already being read by AI systems; they are just not being credited.
- Generate the fan-out. We pull the sub-queries behind each head term, across multiple models and personas, and cross-reference them against the grounding queries Bing reports for your existing pages.
- Cut it down. A single head term will produce three to four hundred raw sub-queries. Most are noise. We consolidate to roughly 20 to 25 primary targets and a short list of secondary ones, then layer on search volume and clustering to find the gaps your competitors are filling and you are not.
- Assign a content type to every surviving cluster. This is where PROOF does the work. Each cluster gets matched to the asset that suits it, rather than everything defaulting to an article.
- Measure, then repeat. AI answers move faster than organic rankings do, which means you find out whether it worked in weeks rather than quarters.
What you get out the other end is a Fan-out Coverage Scorecard: every cluster, what you currently cover, where the gap is, how big the opportunity is and what to build. One page, and it is the thing that turns all of this into a list your team can actually work through.
Our process builds on two people who have done the serious work here: Mike King at iPullRank, on how fan-out actually behaves and how to turn one search into a content plan, and Cyrus Shepard at Zyppy, on the research workflow. We have adapted both for hub building, and we part company with Shepard on the ranking question above.
How we prove it worked
“How would we even know?” is the fair question, and most of this industry answers it badly.
Bing Webmaster Tools’ AI Performance report is currently the most useful free source: it shows how often each URL was cited in AI answers, and the grounding queries that led to those citations. Google Search Console’s Generative AI features report shows how often your pages appeared in AI Mode and AI Overviews, though not the queries behind them. Beyond those, platforms like Profound, Otterly, Peek and Ahrefs Brand Radar sample AI answers at scale to track share of answer, citations and competitor visibility.
None of them is complete on its own. Together they are enough to set a baseline, show movement and report honestly.
Working with us
AI Search & GEO Audit
The full method applied to your site: technical validation, fan-out research across your priority topics, competitor and citation analysis, and a prioritised roadmap split by owner so your developers, your content team and ours each know what they are picking up.
AI Visibility Benchmark
A measured baseline of where you currently appear across AI platforms, who is being cited instead of you, and a repeatable methodology so the next report is comparable to this one. Suited to you if AI visibility is going to be reported on rather than just discussed.
Ongoing optimisation
Building the hubs, briefing and reviewing the content, refreshing the assets that go stale, and tracking movement. This is where the roadmap becomes pages.
Related: Search Engine Optimisation, Digital Strategy, Data & Analytics, Digital Marketing Training.
Where to start
Get the framework. The PROOF guide covers the five stages, the five-step research method and the Fan-out Coverage Scorecard template, written for marketing managers rather than SEOs. Download the guide.
Train your team. AI in Digital Marketing is a six-session program for marketing, content and creative teams working in Gemini or Copilot. See the course outline.
Talk to us. If you want to know where you currently stand before committing to anything, that is what the audit is for. Get in touch.
AI Search & GEO: common questions
Is SEO dead?
No, and be careful of anyone who says it is. Ranking well still correlates with being cited, still drives the traffic AI does not intercept, and still underpins the technical health an AI crawler depends on. What has changed is that ranking alone no longer gets you into the answer. SEO is now the floor rather than the ceiling.
Can you guarantee we will appear in ChatGPT?
No. Nobody can, and a guarantee should be treated as a warning sign. AI answers are probabilistic and personalised, which means two people asking the same question can get different answers on the same day. What we can do is measurably increase how often you are cited, across a large enough sample of queries for the change to be real rather than anecdotal.
We already get cited in AI Overviews. Isn’t that enough?
Not necessarily. Research across 100 B2B category queries in 2026 found that when brands were cited from their own self-promotional listicles, Google went on to recommend a competitor roughly 69% of the time. Being used as a source and being named as the answer are two different outcomes, and only one of them sells anything. It is worth checking which one you are actually getting.
How is this different from what our SEO agency already does?
Three things. We research the sub-queries behind a question rather than the keyword itself. We assign a content type to each one instead of defaulting to articles. And we measure on the surfaces that actually report AI citations, which are not the ones in a standard SEO report.
How long before we see anything?
Faster than traditional SEO, which is one of the few genuinely good pieces of news here. AI answer systems draw on a much shorter window of data than organic rankings do, so a page updated this month can influence an AI answer well before it moves in the blue links.
Do we need to publish more content?
Usually less, not more. Most sites we look at already have the raw material and are spending their budget on volume rather than coverage. The first pass is almost always about restructuring and extending what exists before anything new gets commissioned.
What is a fan-out query?
A sub-question an AI generates for itself, in the background, on the way to answering the question you actually asked. One question typically produces dozens. They are where AI citations are won, and they are invisible in every keyword tool you currently use.
See some of our work
