Generative Engine Optimization (GEO): How to Get Your Product Cited by ChatGPT and Perplexity

For two decades, getting found online meant ranking on Google. That assumption is breaking. A growing share of buyers now start by typing a question into ChatGPT, Perplexity, or Google's AI Overview, and instead of ten links they get one synthesized answer naming a handful of options.

If your product isn't in that answer, you don't exist for that buyer. There is no page two of an AI answer. Generative Engine Optimization (GEO) is the work of getting into it.

This guide is longer than most on the topic because I've rewritten it twice after our own data contradicted things I'd confidently published. I've kept the corrections in rather than quietly editing them out.

What GEO is

Generative Engine Optimization is the practice of getting your brand mentioned and cited inside AI-generated answers. When someone asks "what's the best tool for finding leads on Reddit?", GEO is what determines whether your product is one of the three names the model returns.

The term comes from a paper by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at KDD 2024. They built a benchmark of about 10,000 queries, ran them through a generative engine, and tested nine different edits to source content to see which raised the share of the answer attributed to that source. That paper is why the acronym exists, and it's still the only widely cited controlled study in the space. Nearly everything else, including most of this guide, is observation.

GEO vs SEO

The difference starts with the shape of the result.

The row that changes strategy most is the second. Ranking eighth in Google is a real if modest outcome. Being the model's fourth choice when it names three is nothing at all. That compresses the value of "pretty good" and raises the value of being the single most direct answer to a narrow question.

The row people notice least is the fourth. Search queries are short because twenty years of Google trained us to strip sentences down to keywords. Nobody does that with an assistant. They type the whole situation: "we're a 4-person agency already living in Notion, what CRM makes sense". Your keyword research does not contain that string, which is why keyword tools are close to useless for building a GEO question set.

How an AI answer actually gets built

You can't reason about GEO without a rough model of the pipeline. For a grounded engine, it goes something like:

1. Query understanding and fan-out. Your question becomes one or several search queries. A single buyer question routinely produces three or four internal searches, which is one reason the same question returns different sources on different runs.
2. Retrieval. Each query hits an index and returns candidate documents, already split into chunks of a few hundred words. Matching happens on the chunk.
3. Reranking. A smaller model scores each candidate chunk for how well it answers the question specifically. This stage is where a lot of "why on earth did it pick that page" gets decided, and it optimises for answering rather than for authority.
4. Synthesis. The top chunks go into the model's context with an instruction to answer and attribute. The model writes prose and attaches citations to the parts it drew from.
5. Attribution. What you see as a citation is the model's account of which source it used, produced after the fact. It's usually right and it isn't a log.

Ungrounded answers skip steps 1 to 4 entirely. The model just writes from memory. Both look identical in the chat window, which is the single biggest source of bad GEO analysis, including my own for about six weeks.

The consequence: passages, not pages

Since matching happens on chunks, your page never competes as a page. This explains most of the behaviour that otherwise looks arbitrary:

• A guide covering twelve distinct sub-questions has twelve chances to be retrieved. The same word count spent restating one idea has one.
• A section that only makes sense after reading the two above it is a section the model can't lift, because it arrives without that context.
• Small sites with no domain authority routinely win specific questions, because nothing on the authoritative pages addressed that exact question.

This is also the honest answer to "should my pages be longer". Length itself is not a signal anywhere in the pipeline. Longer pages get cited more when the extra length is extra answered questions. Padding a page with filler makes the retrieved passage worse and can only hurt.

What the evidence supports

From the Princeton study. Of the nine edits tested, the ones that raised visibility meaningfully were adding relevant statistics, adding direct quotations from credible sources, and adding citations to sources. Reported lifts landed in the region of 30 to 40 percent in relative terms for the best-performing methods, and the effect varied by category. Keyword stuffing performed poorly. Edits aimed at making prose more fluent or more authoritative-sounding did close to nothing.

From our own logs. Since June 2026 we've stored the URLs returned by our visibility checks. The usable sample is small: 23 citations from grounded Perplexity checks across six buying questions between 17 June and 9 July 2026. In it, 14 of 23 sat on the domain of a vendor selling in the category being asked about, 7 were listicles, 1 was a Reddit thread, and there were zero citations to G2, Capterra, Wikipedia, or any comparison page. Full breakdown, with every URL, is in the teardown.

That result cuts against the standard advice, including advice in earlier drafts of this page, that your own site is the weakest possible source about you. On the retrieval path the engines cited vendor pages constantly. One small tool, catchintent.com, got cited three times on one question across three of its own pages, because it had written about the thing being asked.

On the rank disconnect. (We publish who "we" is, and the numbers, including the bad ones.) For one customer audit we pulled every page the engines cited across their question set and checked each against Google's top 100 for the surfacing query. 58 of 96 cited pages were not in the top 100 at all. The same gap across 533 URLs is in our citation statistics. Whatever the exact mix of causes, you cannot use Search Console as a proxy for AI visibility, and your best-ranking pages are not automatically your citation engines.

The playbook

1. Build a real question set

Thirty to fifty questions minimum. Six gave us a directional read and not much else; the counts swing hard on a single answer below about thirty. Sources worth mining: your sales call recordings, support tickets, the "what should I use for" threads in your category's subreddit, and your own onboarding survey. Write them as full sentences the way a person would type them, including the constraints (team size, budget, existing stack) that make the question specific.

Freeze the set. A question set you keep editing produces a trend line that measures your editing.

2. Answer one question completely per section

Use the question as the heading, near-verbatim. Put the answer in the first forty words underneath, self-contained enough to make sense if someone read only that block. Then elaborate. This is unnatural to write and it's what makes a passage retrievable.

3. Load in specifics

Numbers, dates, prices with a "checked on" date, named tools, direct quotes with attribution. This is the intervention with the strongest evidence behind it, and it doubles as the thing that makes a passage worth quoting rather than paraphrasing. Vague copy gives a model nothing to lift.

4. Cover the sub-questions on the same page

Every genuinely distinct sub-question you answer is another retrieval surface. This is the version of "write longer" that actually works, and it has a natural stopping point: when you're out of real questions, stop.

5. Earn mentions off your own site

Roundups, comparison articles, community threads, review profiles. I've moved this below on-site work because the data didn't support putting it first, but it's still how you show up for questions where your own domain reads as self-serving, and it's the main way to build the broad association that might survive into a future model's training.

A warning from experience: doing this by posting promotional content yourself is a fast route to being banned. We lost four Reddit accounts learning that (the write-up is here), and our product no longer asks anyone to post.

6. Fix the machine-readable basics

Clean HTML, real headings, an FAQ block, schema markup, an llms.txt if you like, and no critical content that only exists after JavaScript executes. None of this wins you citations on its own. All of it can cost you them.

Measuring GEO

The metrics worth tracking, roughly in order of usefulness:

Share of answers. Across your fixed question set, in what percentage of answers is your brand named at all? This is the headline number. Track it smoothed over several weeks, because week-to-week noise on a 30-question set is large enough to swamp real movement.

Citation count and cited URLs. Which of your pages are actually being drawn from? Usually a surprise, and usually a shorter list than you expect.

Competitor set. Who else gets named alongside you, and who's entering. New entrants in your answers are an early warning that reads better than any traffic metric.

Grounded flag. Record it per answer. Without it you're averaging citations and recall into a number that describes neither.

Position within the answer. Being named first in a list of three is worth more than being named third, though this is softer than it sounds since answer ordering is not stable.

CueScout runs this loop: your question set against the engines on a schedule, real citations captured from the engines' own annotation fields rather than scraped out of the answer text, plus the domain-level view of which sources the answers in your category keep drawing from. For a one-off baseline, the AI Visibility Checker is free and takes a minute. The methodology page documents how we ask, what we extract, and what the method cannot see.

Common mistakes

Treating "AI search" as one channel. Perplexity grounds every query and cites heavily. ChatGPT decides per query whether to search, so the same question yields memory on Monday and citations on Tuesday. AI Overviews sit on Google's index and carry more SEO over than anything else. Averaging them produces mush.

Optimising for the engine you can measure. Perplexity is the easiest to track and among the smallest by usage. Most GEO tooling leans on it for that reason, including ours in its early versions. Be aware of what that biases.

Chasing the training data. You cannot influence what the next model memorises on any timeline that matters to a quarterly plan. Spend on retrieval.

Reading positions off contaminated data. Our own Search Console impressions were roughly 12 percent scraper traffic firing -site: operator queries at positions 1 to 8, which faked a page-one ranking for a blog post that ranked nowhere. Strip the operator queries before you read anything — how to spot them is its own post.

Publishing volume as the strategy. Twenty thin pages answering the same question badly lose to one page answering twenty questions well, because the twenty thin pages are competing with each other for the same retrieval slot.

What's still open

I don't have a clean answer to the question I get asked most, which is whether longer pages get cited more per page when you hold topic coverage constant. The mechanism argument for passage coverage is solid. The isolated effect of length is unmeasured as far as I know, and anyone telling you they have a number for it probably has a correlation.

I also want to redo the 23-URL sample at 50 questions across categories I don't sell into, since a sample drawn entirely from my own market is about as biased as it gets. When that's done I'll publish the counts either way.

For the mechanics underneath all of this, see how AI engines choose brands to recommend. For the demand side of the work, start with demand intelligence.