Generative Engine Optimization (GEO): How to Rank in ChatGPT

The goal of search optimization changed underneath everyone in the last 18 months, and most marketing teams are still optimizing for a scoreboard that pays out less every quarter.

Ranking number one on Google used to be the win. Now an AI Overview often sits above your number one result and answers the question before the searcher scrolls. How often depends on who is counting: Pew Research Center found an AI summary on 18% of the Google searches its US panel ran in March 2025, and Ahrefs saw one on 20.5% of 146 million desktop results pages in September 2025. When one appears, fewer people click: Seer Interactive measured a 61% drop in organic click-through on those queries in November 2025, with a partial recovery in early 2026.

Meanwhile the answer engines kept growing. More than 800 million people use ChatGPT every week, Sam Altman said at OpenAI's DevDay in October 2025, and Perplexity's CEO put its volume at 780 million queries in May 2025. The traffic did not disappear. More of it now happens inside the answer.

Generative Engine Optimization, or GEO, is the discipline of becoming the source those answers quote. It is not a rebrand of SEO. The mechanics are different, the signals are weighted differently, and the unit of victory is no longer the page. It is the passage. This is the playbook we run inside SEO and GEO engagements, and every outside number in it links to its source.

The reflex is to dismiss AI search as low-intent browsing. The data that exists points the other way. It is still thin, mostly single-site case studies, but it is consistent: visitors who arrive from an AI answer tend to convert better than Google organic, because the engine has already pre-qualified them by recommending a business by name.

On that one site, a ChatGPT referral converted about 9 times as often as a Google organic click. Ahrefs saw the same pattern on its own site in 2025: AI search visitors were 0.5% of traffic and 12.1% of signups. Single-site numbers are not a law, and your trade will behave differently. But when an AI engine names you as the answer, the searcher treats it as a recommendation, not a list of 10 options to compare. You are not fighting for a click anymore. You are fighting to be the recommendation.

The unit of victory

Passage-level citability is the whole game

AI engines do not cite pages. They retrieve and lift passages. A retrieval system breaks your content into chunks, scores each chunk against the query, and quotes the ones that answer it best. This means a paragraph buried in a page that ranks number eight can get cited while the number one result gets ignored, because the number one result wrapped its answer in 3 sentences of throat-clearing context that do not survive extraction.

The test is simple. Take any paragraph from your page, paste it into a document with no heading and no surrounding text, and read it cold. Does it make a complete, specific claim that a stranger could quote and attribute to you? If it leans on the sentence before it ("As we saw above," "This is why"), it is not citable. Rewrite it so it stands alone.

The GEO paper by researchers from Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi (first posted 2023, published at KDD 2024) tested which on-page changes lift visibility in generated answers. Adding statistics, citing sources and including quotations were among the strongest methods, and the best of them raised visibility by up to 40% on the authors' benchmark. Lower-ranked pages gained most: citing sources lifted the visibility of a page ranked fifth by 115%. It is a research benchmark, not a live engine, but it matches what we see: passages dense with named numbers and verifiable facts get extracted, vague prose gets skipped.

What a citable passage looks like in practice

• It opens with the claim, not the windup. The first sentence is the quotable one.
• It contains at least one specific number, named tool, or dated benchmark, with its source.
• It resolves inside 2 to 4 sentences. Engines favor self-contained chunks.
• It uses no pronoun that points outside the paragraph. No "this," "that," or "as mentioned" leaning on prior context.

Structured data tells a machine what your content is before it has to guess from the prose, which is why it is often sold as the secret GEO lever. The best evidence says otherwise. Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, against 4,000 matched pages that never did, and found no major change in citations on AI Overviews, AI Mode or ChatGPT. Their conclusion: adding schema only to win AI citations is a bet the data does not support.

Ship it anyway, because it is correct and cheap: FAQPage for question and answer blocks, ItemList for lists, HowTo for step sequences, Person for the author. It keeps your page eligible for rich results and removes any doubt about what the page is. Just do not expect it to do the work of the passage. The listicle advantage is real for a different reason: a well-built list hands the engine pre-chunked, discrete facts, which is the citability rule above, not the markup.

Trust as a signal

E-E-A-T is a retrieval question, not a vanity metric

Experience, expertise, authoritativeness, and trust used to be a fuzzy quality bar. Google's own guidance on helpful content asks publishers to make clear who created a piece, how, and why. An engine that has to decide which of 20 passages to quote has the same question, and a byline with a real name, a real role, and a link to a real author page with a track record is the cheapest answer to it. Anonymous content, or content bylined to a generic "team," gives it nothing to go on.

The fix is operational, not cosmetic. Every article we ship through the blog pipeline carries a named author, a linked bio, and Person schema that names the author's role and organization. Combine that with first-hand specifics in the prose, real numbers with their sources, and mechanics only an operator would know, and the content reads as experience-backed rather than synthesized.

Ahrefs studied 75,000 brands in 2025 to find what correlates with being mentioned in Google's AI Overviews. The result reorders the off-page playbook. Branded web mentions correlated at 0.664. Backlinks correlated at 0.218. The mention correlation is about 3 times as strong as the link metric the SEO industry has spent 20 years chasing. It is a correlation across brands, not proof of cause, and it covers AI Overviews rather than every engine, but the gap is too wide to ignore.

The top 3 correlated factors in that study are all off-site brand signals: branded web mentions at 0.664, branded anchors at 0.527, and branded search volume at 0.392. The mechanism is intuitive once you understand retrieval. AI engines synthesize answers from the sources that most often reference a brand in context. The more places your name appears alongside the topic you want to own, the more the engine treats you as the canonical answer for that topic. A backlink you can buy. A pattern of unprompted mentions across credible sites is much harder to fake.

Stop buying links. Start manufacturing the conditions under which other people mention you. The engine is counting mentions, not measuring PageRank.

Shivam Bindal

The practical reorder: shift budget from link acquisition toward digital PR, podcast appearances, original data that gets referenced, expert contributions to industry roundups, and getting quoted in the publications your buyers already read. Original data is the most efficient mention engine you can build, because a single proprietary number gets cited again and again across the web, each citation a fresh branded mention. The programmatic SEO playbook covers how to build that kind of dataset in the first place.

The highest-leverage move

Get into the third-party best-of lists

Here is the move almost nobody prioritizes correctly. When someone asks ChatGPT or Perplexity for "the best water damage restoration companies in Phoenix" or "top asbestos testing labs near me," the engine does not invent the list. It pulls from sources that already exist. Yext's 2025 analysis of 6.8 million citations across ChatGPT, Gemini and Perplexity found 44% came from brand websites and 42% from listings, with reviews and social at 8%. It is vendor research built on Yext's own client base, but the split is the useful part: your own site is only half the picture.

You cannot write your own way onto somebody else's list. The leverage is in getting included. That is a campaign: identify the listicles and directories already ranking for your target queries, reach the people who run them, and earn a spot with a genuine reason to be there, a license or certification, a response time you can prove, a finished job outcome. One inclusion in a widely cited "best of" article is simultaneously a high-trust third-party citation and a branded mention in exactly the context the engine retrieves.

1. 01
   Map the lists

   Pull the best-of and comparison articles already ranking and getting cited for your top 10 target queries. Tools like Ahrefs Brand Radar and Perplexity citation trackers surface which sources the engines actually quote.

2. 02
   Audit your absence

   For each list, note whether you appear, where, and how you are described. Wrong description is as costly as absence, because the engine can quote the description word for word.

3. 03
   Earn the inclusion

   Reach the author with a concrete, verifiable reason to add or move you up. A unique number, a named customer result, a certification or service area no competitor covers. Make the edit easy to justify.

4. 04
   Reinforce with your own corroboration

   Publish the supporting data on your own site, so when the engine cross-references the claim from the list, your page confirms it.

You cannot improve what you do not measure, and AI visibility is invisible in Google Search Console. Share-of-citation is the new rank tracking: for your target queries, what percentage of AI answers cite you, and how does that compare to competitors, on each engine. Measure the engines separately, because they cite different shapes of source. In the same Yext data, Gemini drew 52.1% of its citations from brand websites, OpenAI's model drew 48.7% from listings, and Perplexity spread its citations across industry directories.

Freshness is part of it. In Ahrefs' 2025 analysis of 17 million citations, the URLs AI assistants cited were 25.7% fresher on average than the URLs in Google's organic results, and ChatGPT was the assistant most likely to cite recent pages. Changing the date without changing the page does nothing. A scheduled refresh of your highest-value pages, updating the numbers, the dates, and the named tools, is real work that engines can see. Run it through your lead intelligence layer so the underlying data stays current without manual research every quarter.

Build the monitoring stack with purpose-built tools, Ahrefs Brand Radar, Semrush AI Visibility, Scrunch AI, Otterly, or similar, and track a weekly share-of-citation number per engine per query cluster. Treat it like a funnel metric, not a vanity dashboard. When a competitor's share climbs, go find which list or mention drove it and respond. GEO is a program you run continuously, the same way the page library treats every channel as a compounding system rather than a one-time project.

GEO is not a mystery. It is a short list of mechanical moves, run in order, measured weekly. Rewrite your top pages so every paragraph stands alone and carries a sourced number. Ship JSON-LD schema matched to each page's shape because it is correct, not because it is magic. Shift off-page budget from links to mentions and original data. Run a campaign to get into the third-party lists and directories for your top queries. And stand up share-of-citation tracking across ChatGPT, Claude, Gemini, and Perplexity so you can see what is working. The FAQ block on a well-built landing page is a good first place to practice: it is already written as questions a buyer asks, answered in complete sentences.