How to get cited when the buyer never clicks

You get cited by writing a page the model can lift: a short answer, a named statistic, a quotation, and a source a buyer can check. Princeton’s GEO study (KDD 2024) found citations, quotations, and statistics lift visibility by up to ~40%. Keyword stuffing does not.

A marketing director asks ChatGPT which studio should run Meta and Google for a brand that already has a site. Three names come back. Yours is not one of them. Your marketing page ranks. The work is on /work. The buyer never opened either.

That is the job now. Not a new definition of search — SEO, AEO, and GEO are three jobs, and this note is only about the third. You write so the generated answer can name you, quote you, and get the facts right when nobody follows the link.

Why does the click keep falling even when you rank?

Ahrefs re-ran 300,000 keywords in December 2025. When an AI Overview is present, average click-through for the top-ranking page is 58% lower than the forecast without that overview.

Update: AI Overviews Reduce Clicks by 58% — Ahrefs, Dec 2025 data

Pew Research Center tracked 68,879 Google searches from 900 U.S. adults in March 2025. Users who saw an AI summary clicked a traditional result on 8% of visits. Users who did not saw a click on 15%. A click on a link inside the summary happened on 1% of those visits.

Google users are less likely to click on links when an AI summary appears — Pew, July 2025

SparkToro, using Similarweb clickstream for January–April 2026, put U.S. zero-click Google search at 68.01%. In 2024 the same series sat at 60.45%.

In 2026, Less than One Third of Google Searches Still Send a Click — SparkToro, June 2026

Semrush’s 200,000-keyword AIO sample found 82% of desktop AI Overviews on keywords under 1,000 monthly searches, and 35% of those desktop queries were questions.

We Studied 200,000 AI Overviews — Semrush

The ranking is still real. The visit is no longer the unit of work.

What did Princeton actually measure?

Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande published GEO: Generative Engine Optimization at KDD 2024 (arXiv 2311.09735). They built GEO-bench — 10,000 queries — and tested nine rewrites. Visibility was not “did we rank.” It was how much of the generated answer was attributable to that source.

The line teams quote is the right one, if you keep the rest of the table.

“We find that including citations, quotations from relevant sources, and statistics can significantly boost source visibility, with an increase of over 40% across various queries.” — Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024

On GEO-bench, the best methods improved Position-Adjusted Word Count by about 41% and Subjective Impression by about 28% versus an unoptimized baseline. On a live Perplexity setup, quotation addition improved Position-Adjusted Word Count by about 22%, and statistics addition improved Subjective Impression by about 37%.

They also measured combinations. Fluency plus statistics was the strongest pair. Cite-sources helped lower-ranked pages more than the page already sitting at position 1 — up to a 115.1% visibility lift for a fifth-ranked source in one split. That is the useful part for a studio that will never out-domain a publisher.

Google’s AI features guide sits next to the paper, not under it. Write for people. No special AI markup. No llms.txt requirement for Search. Foundational SEO still decides what can be retrieved. The Princeton methods decide whether the retrieved page is the one the answer leans on.

GEO: Generative Engine Optimization — arXiv 2311.09735

How do you write a 40–60 word extract a model can lift?

Open the page. Read only the first screen. If a model had to answer from that screen, would the sentence be usable? The extract is not a meta description. It is the answer: the claim, a number with a year and a named source, who it is for, and what happens next.

Pew’s median AI summary in March 2025 was 67 words. Semrush saw desktop AI Overviews average 119 words. Write the short block so the first lift is clean. The rest of the page exists so the longer answer does not invent the rest.

What does a GEO pass look like on a real service page?

This is the marketing block on /services#marketing, as it ships today.

Before — what a model has to invent around

Campaigns that bring inquiries, not just impressions. Meta, Google, and the creative they need. We write, design, and run the work so you can sit with a cost per lead — not a screenshot of reach. Meta and Google Ads. Creative, landing pages, and offers. Tracking you can defend. Weekly reporting in plain language.

Nothing there is false. Almost nothing there is extractable. No number, no source, no quote, no “who this is not for.” Asked which studio should run Meta and Google, the answer has to guess.

After — same page, Princeton methods applied

Simplex Digital runs Meta and Google Ads as one brief with the landing page: creative, offers, and tracking you can defend in a Monday meeting. Ahrefs (Dec 2025) found AI Overviews correlate with a 58% lower CTR for position 1, so the campaign cannot assume the organic click will arrive. We report cost per inquiry, not reach. This is not a media-only retainer and not a template landing page.

“The inquiry is the only shared KPI that does not lie.” That is the measurement rule. Weekly reporting is a conversation against last week’s inquiries — the same number we sit with on the Monday audit.

Pew (July 2025) found users click a traditional result on 8% of visits when a summary appears, versus 15% when it does not. SparkToro (June 2026) put U.S. zero-click Google search at 68.01% in Jan–Apr 2026. Paid and the page have to close the people who do click.

What the answer can now say

Simplex Digital is a studio that runs Meta and Google with the landing page, and reports cost per inquiry rather than reach. They cite Ahrefs’ December 2025 finding that AI Overviews correlate with a 58% lower CTR for position 1, and they treat the campaign as the close when organic clicks do not arrive.

That is a 52-word lift. Specific enough to be wrong if we are lying, which is the point.

A page that cannot be quoted without becoming generic will not be quoted.

What does a GEO pass look like on a real comparison page?

Service pages get the budget. Comparison pages get the prompt. Buyers type “n8n vs Zapier vs Make for a marketing team,” and they ask ChatGPT the same sentence. Here is the GEO pass on that draft.

Before — the page every agency already wrote

n8n, Zapier, and Make all connect your tools. Zapier is the easiest. Make is visual. n8n is for technical teams. Choose the one that fits your stack and your budget. We can implement any of them.

A model can rewrite that from memory. There is no reason to cite you.

After — statistics, sources, quotes, and a decision rule

For a marketing team that already has a site, ads, and a CRM, pick the runner by who will own the workflow in 90 days — not by the logo. Zapier is the hosted connector with the largest prebuilt catalogue; default when the operator will not touch a server. Make is a visual scenario builder; default when the operator thinks in branches. n8n is source-available, self-hosted or in n8n cloud; default when the team needs code nodes, their own credentials, and a log they can inspect.

Google’s AI features guide (updated July 2026): write for people, do not invent special AI markup, do not treat a new file as a ranking factor. “A chatbot answers. An agent takes action.” If the workflow only drafts an email, any of the three will do. If it has to open the ticket and page a person, the runner that holds credentials and a retry log is the one that matters. We implement n8n when we are also building the site and the agent, so one team owns the failure.

No invented price. The page now has a decision rule, a named source, a quotation, and a “who we are not.”

What the answer can now say

For a marketing team, Zapier fits operators who will not host anything, Make fits visual branching, and n8n fits teams that want a source-available workflow they can inspect. Simplex Digital implements n8n when they are also building the site and the agent, so one team owns the failure.

Forty-nine words. A comparison that refuses to pick a winner is a brochure. A comparison that states the rule is a source.

What should you refuse to do?

Do not stuff the query. Keyword stuffing underperformed the baseline. Google says the same: do not write a page for every fan-out, and do not chase inauthentic mentions.

Do not rewrite the paragraph into “AI-friendly” fragments. Google says there is no requirement to chunk content. A heading and a table beat a pile of 40-word cards.

Do not add a schema type you invented for “GEO.” Structured data still helps rich results. It is not a secret citation tag. The file stack is a map for agents, not a ranking lever.

Do not hide the number behind a login, a PDF, or a demo wall. Retrieval cannot quote what it cannot fetch.

Do not confuse being mentioned with being chosen. G2’s 2026 buyer work, which we treat at length in B2B buyers start in ChatGPT, found 85% of buyers think more highly of a vendor the chatbot includes. The citation is the shortlist. The site still has to close.

How do you check whether you were cited?

You do not need a vendor dashboard. You need the questions the buyer actually types. We ran twelve of those across four engines. That log is four models, twelve questions.

1. Write the twelve questions a buyer would ask before they speak to you. Use their words. “Who should run Meta and Google if the site already exists.” “n8n vs Zapier vs Make for a team of four.”
2. Ask ChatGPT, Perplexity, Gemini, and Google AI Mode the same twelve, on a logged-out or fresh session, in the same week.
3. Record four things: were you named, were you described accurately, which URL was cited, and who else appeared.
4. For every miss, open the page that should have been the source. If the 40–60 word extract is missing, write it. If the statistic has no year, add the year. If the quote is unattributed, attribute it.
5. Repeat the twelve in 30 days. The metric is names and accuracy, not sessions.

Search Console’s generative AI report will not tell you what ChatGPT said. That is why the twelve questions stay in a sheet.

What belongs on the page besides the extract?

The extract gets you into the answer. The rest of the page stops the answer from inventing you.

Put the proof where a crawler can read it. Work is four cases with a brief and a constraint. Eternal is a maison site. Aurelia is a support agent. Meridian is a campaign board. Vesper is a night line. A model that cites us and then describes a SaaS dashboard has misread you.

Put the engagement shape in prose. Discovery is a written note. Build is the site, the campaign, the agent, the workflows. Partner is the review after launch. We write that file properly in the llms.txt note. The HTML page still has to say it.

Put the next step in a sentence a stranger can finish. “Send the current ads account and the last four weeks of landing-page screenshots. We will write back with what we would turn off, and what we would measure this week.” That is a close.

Google’s guide, updated 10 July 2026, says unique, useful content will influence generative visibility more than any other suggestion there. For a studio, that is the case, the weekly number, and the decision rule.

FAQ

How do I get cited by ChatGPT if nobody clicks my site?

You get cited by publishing a page the model can retrieve and quote: a short direct answer, a statistic with a year and a named source, and a quotation with an attribution. Princeton’s GEO paper (KDD 2024 / arXiv 2311.09735) measured lifts of up to about 40% from those three additions. The click is not the mechanism. The extractable paragraph is.

Do Princeton GEO methods still work if I already rank number one?

They work, but the paper’s bigger lifts were on lower-ranked sources. Cite-sources raised visibility for a fifth-ranked page by 115.1% in one split, while the top result sometimes lost share. If you already rank, the job is accuracy: the answer should describe you the way the page does. If you do not rank, the same methods are how a smaller domain gets into the paragraph at all.

Is keyword stuffing useful for AI Overviews or ChatGPT?

No. On GEO-bench, keyword stuffing scored 17.7 on Position-Adjusted Word Count against a 19.3 baseline — worse than leaving the page alone. Google’s AI features guide also tells you not to manufacture a page for every query variation. Repeat the claim once, attach a source, and stop. The model already knows the synonyms.

How long should the citable paragraph be?

Write a 40–60 word block that a person would accept as the answer, then let the rest of the page support it. Pew’s median Google AI summary in March 2025 was 67 words. Semrush saw desktop AI Overviews average 119 words. The short block is for the first lift. The long page is so a deeper answer does not invent your offer.

Should I create a separate “GEO page” for every service?

No. Rewrite the page the buyer is already trying to find. A service page and a comparison page cover most prompts. Extra URLs written only to catch fan-out queries are the scaled-content problem Google already describes. One marketing page that states the number, the source, and the “not for” will outperform five paraphrases of the same heading.

What is the difference between a citation and a mention?

A mention is your name in the sentence. A citation is a name plus a fact the engine can point at — a URL, a statistic, a quote. G2’s 2026 buyer survey found 85% of buyers think more highly of a vendor the chatbot includes. A mention without a fact is easy to drop. A citation with a number is harder to replace.

Do I need llms.txt to get cited?

No. Google’s guide says Search does not use llms.txt, and that creating one neither helps nor harms rankings there. The file is still useful as a map for agents that choose to read it — pricing, work, and the pages you want quoted. Build the extract on the HTML page first. Add the file stack second. We wrote that order up in llms.txt and the file stack.

How often should we re-ask the twelve buyer questions?

Every 30 days is enough to see whether a rewrite moved a name, and not so often that you chase noise. Use the same wording, the same four engines, and a fresh session. Record names, accuracy, cited URLs, and who else appeared. If the description is wrong, the page is wrong. If you are absent, the extract is missing or the page is not being retrieved.

Send the URL you think should be the source — a service page or a comparison — and the twelve questions a buyer would ask before they speak to you. We will write back with the 40–60 word extract as it stands, the three lines we would add, and the claims we would cut. Keep the note either way. Contact the studio.