An AI assistant answers a question by searching the web, reading a small number of pages and writing one answer that credits a few of them. Every step of that process is a filter. A page that is not crawlable and ranking is never retrieved, a page without a clear passage is read but not quoted, and a brand described inconsistently across the web is quoted but not named. This article sets out how the selection actually works, what the click data shows about why it matters, and the four stages most companies move through on the way from invisible to recommended.
What the click data actually shows
The Pew Research Center tracked the real browsing of 900 US adults through March 2025 and published the results in July 2025. The findings are specific enough to plan around.
| What Pew measured | With an AI summary | Without one |
|---|---|---|
| Visits where the user clicked a traditional search result | 8% | 15% |
| Visits where the user clicked a source inside the summary | 1% | Not applicable |
| Pages where the user ended the browsing session | 26% | 16% |
Three more numbers from the same study shape how a page should be written. An AI summary appeared on about 18% of the Google searches in the study. 88% of those summaries cited three or more sources. The median summary was 67 words long, with a range from seven to 369.
Read together, those last two figures describe the opportunity precisely. Most answers have several citation slots rather than one, so this is not a winner-takes-all contest. But the answer itself is short, so the passage that gets absorbed into roughly 70 words has to be compact and self-contained before it can be used at all.
The direction was forecast earlier. In February 2024 Gartner predicted that traditional search engine volume would drop 25% by 2026 as people moved to AI chatbots and virtual agents. Whether the final figure lands at 25% matters less than the pattern the Pew data already confirms: the answer is increasingly consumed in place, and the click that used to follow it is optional.
How an assistant picks the sources it cites
Most AI search products use retrieval-augmented generation. The assistant rewrites the question into search queries, retrieves candidate pages, selects passages from them, and writes an answer that credits what it used. Each step asks something different of your page.
| Step | What the assistant does | What your page needs |
|---|---|---|
| 1. Interpret | Works out what is being asked: a definition, a number, a comparison or a recommendation | A page aimed at one clear question, with headings worded the way people ask it |
| 2. Retrieve | Searches and selects a small set of pages it considers relevant and trustworthy | To be crawlable, indexed and ranking. Without this, nothing else matters |
| 3. Select and synthesise | Reads the candidates and lifts the passages that answer the question | Short sections that answer directly, with specific figures and explicit nouns |
| 4. Cite | Credits the sources it relied on, usually as numbered links | Consistent brand facts, named authors and corroboration elsewhere on the web |
The mechanics are covered in more depth in retrieval-augmented generation and query fan-out, and the passage-level behaviour in context windows and token budgets.
Why SEO still decides whether you are in the running
Step two is the one teams skip. GEO is often sold as a replacement for SEO, and the retrieval step makes that impossible: if a page is blocked, unindexed or ranking on page four, the assistant never reads it, however good the content is. The common failures are mundane. Content that only appears after JavaScript runs is invisible to crawlers that do not execute it, which is most AI crawlers. A firewall rule can block an AI search crawler while leaving Googlebot alone.
Two practical checks come first: whether AI crawlers can reach your pages at all, and whether your framework serves content in the server HTML. If a CDN sits in front of the site, check what it is turning away. The relationship between the disciplines is set out in GEO vs SEO.
The four stages of AI visibility
Visibility arrives in stages rather than all at once, and the work that moves you up changes at each one. Describing it as four stages is Brilad’s framing, not an industry standard, but it matches what we see when we run baselines for new clients.
| Stage | What you see in AI answers | What moves you up |
|---|---|---|
| 0. Absent | Your question gets answered, and competitors are named. Your company appears nowhere | Crawler access, indexing fixes, consistent company facts, a first few third-party mentions |
| 1. Cited | One of your pages shows up as a numbered source for a specific fact | Answer-first rewrites, your own figures and methods, more pages that can be retrieved |
| 2. Recommended | Asked who to use, the assistant names your company among its suggestions | Named experts, trade coverage, reviews and directories, client evidence |
| 3. Described as a leader | The assistant characterises you as a go-to source in your category | Original data nobody else has, and the same description of you repeated across the web |
The hardest move is from absent to cited, because you begin with no trust signals and every fix has to be made at once. Going from cited to recommended is usually faster: by then the technical base holds, the team understands what a quotable passage looks like, and corroboration accumulates. If you do not know your current stage, start by measuring it rather than guessing.
What the research says actually works
The founding academic work on this is “GEO: Generative Engine Optimization” by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at KDD 2024. The authors built a benchmark of varied user queries, applied different content changes, and measured how often generative engines used the result. Their headline finding is that the best methods raised visibility in generated answers by up to about 40%. The methods that did it were adding statistics, adding quotations from credible sources, and citing authoritative references. Keyword stuffing, the old reflex, gave little or no benefit.
Two honest caveats. The study measured visibility inside generated answers under test conditions, not revenue, and the size of the effect varied by subject area, which is why we test changes on a client’s own questions instead of applying a fixed checklist. Treat 40% as evidence that the direction is right, not as a number to forecast with.
| What the research rewards | What that means on a page |
|---|---|
| Statistics | A figure you measured, with the period and the sample stated next to it |
| Quotations | A short comment from a named person with a real role, not an anonymous “our experts say” |
| Cited sources | Primary references a reader can open and check |
| Clarity | Short sentences, explicit nouns rather than pronouns, one idea per paragraph |
Why this matters more than it sounds: a figure that appears on ten sites can be credited to any of them, while a figure only you hold has one possible source. That argument is developed in information gain and in why structure alone is no longer a differentiator.
The work, in the order it pays off
Doing this in the wrong order wastes money. Rewriting pages before fixing crawler access means the rewrites are never read. Four phases, each ending with something shipped and something measured:
Phase 1: measure where you actually are
- Write 15 to 20 questions a real buyer would ask, in their words, not your product names
- Run each one on ChatGPT, Gemini, Perplexity and Google AI Overviews, several times, and record what comes back
- Search your brand name while excluding your own domain, to see your third-party footprint
- Keep the raw answers. This is the baseline everything later is compared against
Phase 2: fix access and identity
- Confirm each AI crawler can fetch your key templates, and that content is in the server HTML
- Make the company name, founding year, location and description identical on your site, LinkedIn and every directory
- Publish one authoritative About page with Organization schema and sameAs links to your real profiles
- Give the people who write your content author pages with their actual experience
Phase 3: make the pages extractable
- Rewrite your highest-value pages answer-first, with the figure in the opening lines
- Turn section headings into the questions buyers actually ask
- Add a methodology note and a visible last-updated date wherever you publish numbers
- Check that nothing important hides behind tabs or accordions that load on click
Phase 4: build corroboration and measure again
- Place two data stories with trade media or industry associations
- Get listed, accurately, on the review sites assistants cite
- Re-run the same question set and compare against the baseline, not against memory
- Set a monthly cadence so the trend is visible rather than anecdotal
For the publishing calendar that sits alongside this, see our 90-day AEO and GEO content plan. For realistic timings, how long GEO takes to show results sets out what moves in weeks and what takes months.
The mistakes that keep teams at stage zero
- Treating GEO as a replacement for SEO, then wondering why rewritten pages are never quoted
- Burying the answer under three paragraphs of introduction, so a competitor’s cleaner passage gets lifted instead
- Letting company facts drift apart across the website, LinkedIn and directories, which leaves the assistant unsure who you are
- Publishing figures with no date and no method, which makes them unattributable and easy to skip
- Measuring traffic only, and so missing citations and recommendations that never produce a click
- Reacting to every platform announcement instead of fixing the foundations that hold across all of them
Frequently asked questions
Does GEO replace SEO?
No. Assistants retrieve pages from a search index before they answer, so a page that is not crawlable, indexed and ranking is rarely read. SEO decides whether you are a candidate, and the rest decides whether you are quoted.
Is it too late to start in 2026?
No. Most companies have not done any of this deliberately, and many brands that appear in AI answers today got there through years of accumulated authority rather than a plan. Deliberate work compounds faster because the fixes reinforce each other: access makes rewrites readable, and consistent identity makes citations turn into recommendations.
What is the difference between a citation and a recommendation?
A citation means one of your pages is listed as a source, which shows the assistant could retrieve and use you. A recommendation means it suggests your company when asked who to use, which requires that it be confident about what you are. Recommendations generally need stronger identity signals and more third-party corroboration.
Can a small company compete with a large one here?
On specific questions, often yes. Assistants quote passages, so a small company with clear, dated, well-sourced detail on a narrow subject can be cited ahead of a larger one whose pages are hard to extract. Consistent company facts and a few targeted trade mentions close more of the gap than budget does.
Can anyone guarantee that an assistant will recommend us?
No. Nobody outside the AI companies controls their answers, and a guaranteed placement offer is a warning sign. What can be done is to make you the clearest available source, then measure mention rate on a fixed question set over months.
Which platform should we start with?
Whichever your buyers use, which the baseline tells you. The underlying work is largely shared: access, identity and extractable passages help everywhere. Platform-specific guides are available for ChatGPT search, Perplexity, Gemini and Copilot.
Sources
- Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, 22 July 2025, based on the tracked browsing of 900 US adults during March 2025
- Gartner, “Gartner Predicts Search Engine Volume Will Drop 25% by 2026 Due to AI Chatbots and Other Virtual Agents”, 19 February 2024
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, “GEO: Generative Engine Optimization”, KDD 2024
- Google Search Central documentation on AI features in Search, including query fan-out