Getting cited when nobody clicks: the GEO playbook for pages that don't rank first

In the first four months of 2026, 68.01% of US Google searches ended without a click. That is SparkToro's number, calculated from Similarweb clickstream data, and it is up from a restated 60.45% in 2024. SparkToro is careful that the two years ran on different panels, so read the 7.56-point gap as a direction rather than a measurement. The series is genuinely bumpy on the raw numbers — its 2021 reading came in higher than its 2024 one — but every version of it has landed far above where this sat a decade ago.

Here is the part that stings, and it predates AI answers being any good: back in 2024, SparkToro counted 360 open-web clicks for every 1,000 US Google searches. Even then, roughly two thirds of US searches produced no click to anyone else's site.

The ranking you have been optimising for increasingly decides who gets summarised, not who gets visited.

Most advice about this is vague. One piece of research gets specific, and one of its findings changes who should care most.

The finding worth acting on tonight

The founding paper is "GEO: Generative Engine Optimization" (Aggarwal et al., KDD 2024, arXiv 2311.09735). Everyone quotes its headline: optimising content for generative engines can lift visibility by up to 40%.

Almost nobody quotes the next part. The gains run backwards:

"the Cite Sources method led to a substantial 115.1% increase in visibility for websites ranked fifth in SERP, while on average, the visibility of the top-ranked website decreased by 30.3%."

Read that twice. The same edit, adding inline citations to your own claims, more than doubled AI visibility for a page sitting fifth. It also cost the page sitting first almost a third of its visibility.

This is the one thing to take away, and you can act on it without signing up for anything: generative engine optimisation is a tactic for the site that is losing. If you already rank first, the AI already has you. Dressing your page in citations mostly dilutes what made you the obvious source. If you rank fifth, you are invisible in the summary regardless of the click you were never going to get. The same edit is the cheapest visibility you will buy this quarter.

That flips the usual order of operations. Traditional SEO rewards the incumbent: authority compounds, backlinks accrue, and the top result stays top. This does the opposite. It is the rare channel where being the smaller site is the qualification.

A sample, so you can see the shape of the work

The checklist below has twelve points. Here is the one that costs nothing and gets skipped most often because it feels like a writing preference instead of a mechanism:

The citation anchor test. Read every paragraph and ask: could this single paragraph be pulled into an AI answer and still make sense with nothing around it? If yes, it is a citation anchor. Every section should have at least one, and they belong at the start of paragraphs, where extraction happens.

A model assembling an answer looks for a self-contained unit it can lift. A sentence that only makes sense with the three paragraphs above it cannot be lifted, so it is left behind. "It depends on your stage and your market" survives no extraction at all. "Raise when you have six to twelve months of runway and a clear next milestone" survives every extraction.

That is why hedging is expensive here. The engine cites whoever was willing to answer.

Below you will find the 12-point pre-publish checklist, four content templates built for extraction, and a free 30-minute weekly audit that tells you whether any of it is working. It also includes five mistakes worth checking a draft against and a five-day sprint that needs no new tooling.

Artifact 1 — The 12-point AEO checklist

Run this on a draft before it ships. The list is ordered by impact.

Structure (highest impact)

• [ ] 1. Lead with the answer. Your main point goes in the first 40–50 words. AI answers and featured snippets both pull from the opening. Bury the insight three paragraphs down and the engine takes it from whoever led with it.
• [ ] 2. Use question-based headings. Write H2s as questions people actually type. "How much does a seed round cost?" gets pulled. "Seed Round Considerations" does not. Match the phrasing someone uses with a chatbot, which differs from the phrasing they used with Google.
• [ ] 3. Define your terms explicitly. Put one sentence in the shape of [Term] is [definition] near the top. A definition written as its own sentence is the easiest thing on the page to lift whole, which is the entire mechanism.
• [ ] 4. Write a real FAQ section; add the schema if it is cheap. Google retired FAQ rich results on 7 May 2026, so the markup buys you nothing in the SERP any more, and Google's own generative-AI search guidance says structured data is not required for AI Overviews or AI Mode. No AI vendor has confirmed that FAQ markup improves citation odds — treat that claim as unproven. What is not in doubt is the shape: a page carrying explicit questions with self-contained answers is the most liftable thing you can publish, markup or no markup. Write the section for that reason. If your CMS adds FAQPage JSON-LD for free, leave it on; it is still a valid schema.org type and costs nothing. Budget about fifteen minutes a page if you are hand-writing it.

Content (the real differentiators)

• [ ] 5. Write three to five citation anchors per piece. These are standalone sentences that survive removal from their context. Place them at the start of paragraphs. See the test above.
• [ ] 6. Include original data. Run a survey, analyse public data, or interview practitioners and aggregate what they say. If your page says what fifty other pages say, there is no reason to cite yours in particular.
• [ ] 7. Build and name your own frameworks. A named concept is something an engine can reference by name. That is more durable than being one source among many. This one rests on mechanism, not on the GEO paper — the paper's nearest method, adding unique terminology, sits in its explicitly non-performing bucket.
• [ ] 8. Update on a schedule. Carry a visible "last updated" date, refresh figures annually, and move examples to the current year. Put a quarterly reminder against your ten best pages.

Technical (the foundation)

• [ ] 9. Implement schema markup. Article, FAQ, HowTo, Organization. Validate at validator.schema.org, not Google's Rich Results Test — that tool only reports types still eligible for a rich result, which no longer includes FAQ or HowTo. On WordPress, a plugin covers most of it.
• [ ] 10. Keep the page fast. Retrieval runs on a budget, and a page slow enough to time out is a page that never gets read. Three seconds is the usual working target, borrowed from general web-performance practice rather than from any published crawler limit.
• [ ] 11. Build a few good backlinks. One link from a publication your audience respects is worth more than fifty directory links.
• [ ] 12. Check mobile. Googlebot has crawled mobile-first since 2024, so a broken mobile render costs you in Google directly. The AI crawlers (GPTBot, PerplexityBot and the rest) do not advertise a mobile-first rule — for them the thing that matters is that your content is in the HTML rather than assembled by JavaScript after load.

Artifact 2 — Four templates built for extraction

Pick the one closest to how you already write.

Template 1: The definitive answer

1. The question as your H1
2. The answer in the first 40–50 words
3. Two or three sentences of supporting evidence
4. A bulleted list of the key points
5. The expanded explanation
6. An FAQ section covering the adjacent questions

Worked example: "What is product-market fit?" Start with the definition, cover the signs you have it and how to measure it, then add the FAQ. When the H1 matches the query and the first paragraph delivers, you become the default source for that question.

Template 2: The framework article

1. The framework's name as your H1
2. One sentence on the problem it solves
3. The framework, step by step
4. A visual or bulleted breakdown
5. The framework applied to named companies
6. When to use it, and when not to

Worked example: "The ICE Framework for Prioritisation" should cover what it is, how to score each dimension, real team examples, and its honest limits.

Template 3: The data-driven insight

1. The finding as your H1 (a specific number, not a topic)
2. The data that proves it
3. The methodology, briefly
4. What it means for the reader
5. What to do about it

This is the strongest of the four for one structural reason: proprietary data exists only in your content. An engine can paraphrase common knowledge from anywhere and owes nobody a citation. It cannot paraphrase your number without you.

Template 4: The comparison

1. "X vs Y" as your H1
2. A summary table, immediately
3. When to choose X
4. When to choose Y
5. The detailed breakdown, dimension by dimension
6. A clear recommendation

"Which is better" is one of the most common question shapes people bring to a chatbot. The summary table is the most extractable object on the page. Structured rows survive lifting in a way that prose does not.

Artifact 3 — Measuring it, starting free

The 30-minute weekly audit

1. List your ten most important topics or queries.
2. Ask each one of ChatGPT, Claude and Perplexity.
3. For every query, record four things: were you cited by name, were you paraphrased without attribution, were you absent entirely, and which competitors were cited instead.
4. Keep it in a spreadsheet, one row per query per platform.

Four weeks of that gives you a baseline and a gap map. Keep it weekly after that — the same ten queries every time, because citation patterns move when models update and a changing question set tells you nothing.

What to track monthly

1. Citation count across your audited queries, split by platform
2. Brand mentions, split three ways: cited with a link, named without a link, paraphrased with no credit
3. Referral traffic from AI sources (watch for perplexity.ai in your analytics)
4. Featured snippet ownership on your target queries
5. "People Also Ask" appearances on your core topics

Tooling. Google Search Console and your own analytics are free and cover the essentials — Perplexity hands site owners no referral dashboard, so the perplexity.ai referrer in your analytics is the read you get. Paid trackers exist (Otterly.ai, Profound, BrightEdge among them). The category is young enough that manual auditing plus analytics is a defensible position for a small team. Do not buy a tracker before you have four weeks of manual baseline. You will not know what it is telling you.

Artifact 4 — The publishing loop

Step 1: Query research, before you write. Ask your target question of ChatGPT, Claude and Perplexity. Note which sources they cite. Look for the gap: the angle nobody covers, the number nobody has, or the framework that would make the answer genuinely better. When an engine has an incomplete answer to a common question, whoever fills that gap first becomes the default citation.

Step 2: Write for extraction. Put the answer block in the opening and citation anchors throughout. Add original data or a named framework. Keep definitions near the top. Finish with an FAQ section matching the questions people actually ask.

Step 3: Technical pass, after publishing. Add schema markup and validate it at validator.schema.org. Keep the page fast (see item 10), submit it to Search Console, and get a backlink or two from somewhere credible.

Step 4: Monitor and iterate, monthly. Check citations, note what worked, rewrite what underperformed, and do more of the format that got pulled.

Artifact 5 — Five ways to stay invisible

Use this as a red-flag pass on a draft. If any of these apply, fix them before shipping.

1. Optimising for ranking only. You can hold position one and be absent from every AI answer about your own topic. Ranking and being cited are separate outcomes with separate mechanics.
2. Hedging. "It depends" is unciteable. The engine quotes whoever committed to an answer.
3. No original data. This is the most common gap between the companies that get cited and the ones that do not.
4. Stale pages. A guide dated two years ago with figures to match loses on any current topic.
5. No measurement. You cannot improve a channel you are not watching. The weekly audit takes thirty minutes.

Artifact 6 — The five-day sprint

No new tools required.

Day 1: Audit. Run your top five pages against the 12-point checklist. Most sites fail the same three: question-based headings, explicit definitions, FAQ schema. Fix those immediately.

Day 2: Baseline. Run the Artifact 3 audit once, exactly as written — all three engines, all four fields, including which competitors got cited instead. Use the same ten queries you will use every week after this. A baseline recorded differently from the weekly audit is not a baseline; it is a one-off you cannot compare anything to.

Day 3: Rewrite. Take one high-potential piece and rebuild it on Template 1 or Template 2. Give it an answer block, citation anchors, and an FAQ section.

Day 4: FAQ sections. Add an explicit question-and-answer section to your three highest-traffic pages — real questions, self-contained answers. That is the part that gets lifted. If Rank Math or Yoast is already emitting FAQPage JSON-LD on WordPress, leave it on; hand-writing it costs about fifteen minutes a page and buys you nothing in the SERP, so do it only if it is easy. Validate anything you do add at validator.schema.org.

Day 5: Measurement. Build the tracking sheet with the five monthly metrics. Set the recurring reminder for the weekly audit.

At the end of the week, you have a baseline, one properly optimised piece, schema on your best pages, and a measurement habit. That is further than most companies have gone.

The one thing to remember

Go back to the rank-five result. 115.1% up for the page sitting fifth, 30.3% down for the page sitting first. Every artifact above is worth more to you when you currently win less. That is the opposite of how every other distribution channel has ever worked. If you are the incumbent on your topic, run the checklist and measure carefully because some of this can cost you. If you are not, this is the cheapest visibility available to you right now.

Fill in your own baseline before you change anything. Otherwise, in three months you will have opinions instead of a measurement.