GENERATIVE ENGINE
OPTIMIZATION

What is generative engine optimization?

Generative engine optimization (GEO) is the practice of getting your brand named, and your pages cited, inside answers written by AI engines: ChatGPT, Google AI Overviews and AI Mode, Gemini, Perplexity and Claude. It optimizes for retrieval and synthesis rather than for a ranking position, and is measured by mention rate across real buyer questions.

Why GEO exists

The thing that broke the old model is not that AI answers are better. It is that they end the session. The Pew Research Center tracked 68,879 real Google searches from 900 US adults and found that when an AI summary appeared, people clicked a traditional search result on 8% of visits, against 15% of visits without one. Clicking a link inside the summary happened on 1% of visits (Pew Research Center, July 2025).

Read that second number again, because it sets the whole strategy. Being the source under an AI answer is worth almost nothing in clicks. What is worth something is being named in the answer itself, because that is what the reader takes away and searches for later. GEO that chases referral traffic is optimizing the 1%. GEO that chases mentions is optimizing the recommendation.

The traffic that does arrive is unusually good, which is the other half of the case. Ahrefs published 30 days of its own analytics showing AI search at 0.5% of visits but 12.1% of signups, roughly a 23x conversion rate against organic search (Ahrefs, June 2025). The mechanism is not mysterious: the engine did the comparison shopping before the click, so whoever arrives has already been pre-sold. Semrush puts the average AI visit at about 4.4x the value of an organic one and projects AI search visitors overtaking traditional search visitors around 2028.

So the opportunity is asymmetric. AI referrals are still low single-digit percentages of most sites' traffic, which is exactly why your competitors have not staffed this yet, and the visitors are worth several times what your existing organic visitors are worth. That gap is the entire argument for doing GEO now rather than in two years.

Generative engine optimization vs. SEO

The honest answer to "is GEO just SEO again?" is: about 80% of it is, and the other 20% is genuinely new. Crawlability, server-side rendering, clean information architecture, topical depth and being cited by other people are the same levers they always were. Google's own documentation is blunt about the overlap: its guidance for AI features says to focus on "creating helpful, reliable, people-first content" and that there is no special markup you need (Google Search Central).

What is new is the unit of competition. A ranking is a list position, so ten results fit on a page and being sixth still gets you seen. An AI answer is a paragraph, so three or four brands fit and there is no page two. That changes the target from "rank for a keyword" to "be one of the handful of entities the model associates with this category", which is closer to brand and PR work than to on-page optimization.

How does GEO work?

You cannot optimize a process you have not looked at. Almost every bad GEO decision comes from imagining the model is answering from memory, when for anything commercial it is usually running a search and reading the results. Those are two different systems with two different levers.

The model's training data is frozen and enormous, and you cannot meaningfully edit it. What it does hold is a rough prior about which brands belong in which category, built from years of the open web, and that prior is why a model will confidently list your competitor before searching anything. The retrieval layer is live, and it is where the work happens: the engine fires searches, fetches pages, and writes an answer grounded in what came back. Retrieval is winnable this quarter. The prior is a two-year brand project.

GEO best practices that have evidence behind them

Most GEO advice is assertion. There are two pieces of work worth building a content standard on, and between them they explain most of the wins we see.

The first is the paper that named the field. Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024), built a 10,000-query benchmark and tested nine content edits to see which raised a source's visibility in generated answers (arXiv:2311.09735). The headline is a lift of up to 40%, but the useful part is the ranking. The three best edits were adding quotations from credible people, adding statistics in place of vague claims, and citing sources. The worst of the nine was keyword stuffing. The pattern is that what wins is the stuff that makes a passage safe to repeat: an engine that has to stand behind an answer prefers material that is already attributed.

The second is positional. Kevin Indig's analysis of roughly 1.2 million ChatGPT answers found 44.2% of citations were drawn from the first 30% of a page's content, with the remainder split across the middle and the end. The practical reading is not "write short pages". It is that burying your answer six paragraphs into a warm-up costs you the citation, and moving it to the top is usually a one-hour edit rather than a rewrite.

Put those together and you get a content standard with evidence behind it: answer in the opening, attribute your numbers, quote named people, and stop padding. Note how much of that is also just good editing. GEO techniques that would embarrass you in front of a human editor are generally the ones that do not work.

The technical layer

Technical GEO is a short list and the tips below are nearly all prerequisite work: it does not win you mentions, it stops you being invisible. Do it once, verify it, move on to content. The single most common own-goal is a robots.txt rule written to block AI training that also blocks AI search.

Those are different crawlers, and the distinction is worth learning properly because it is where good intentions cost real visibility. Blocking GPTBot opts you out of OpenAI's model training and does nothing to ChatGPT's search citations. Blocking OAI-SearchBot removes you from ChatGPT search. Blocking Google-Extended opts out of Gemini grounding and training, and has no effect at all on AI Overviews or AI Mode, which run on Googlebot. If you block Googlebot to keep out of AI Overviews, you have also left Google Search.

Technical GEO best practices: four checks

Do you need an llms.txt file?

Short answer: it will not help you in Google, and most GEO guides that recommend it have not checked. This is worth being precise about, because it is the single most over-prescribed tactic in the field.

Google has said no, repeatedly and on the record. Gary Illyes confirmed at Search Central Live that Google does not support llms.txt and has no plans to. John Mueller compared it to the keywords meta tag, which is a pointed comparison: a file where the site owner declares what the site is about is exactly the kind of self-report search engines learned to ignore decades ago. Google's AI features documentation now states it plainly: "You don't need to create new machine readable files, AI text files, or markup to appear in these features" (Google Search Central).

So why does CITED's own action plan still suggest publishing one? Because the cost is an afternoon and the downside is zero, some non-Google tooling does read it, and the exercise itself is genuinely useful: writing down the ten pages you want quoted, each with a sentence on what it covers, forces the prioritisation most teams have never done. Treat it as free hygiene and a planning artifact. Do not treat it as a lever, do not put it in phase one, and be sceptical of any GEO vendor whose recommendation list opens with it.

How to optimize content for GEO: writing that gets cited

This is where GEO is actually won, and the rules follow directly from the mechanics: engines lift passages, they prefer attributed material, and they read the top of the page hardest.

One page, one question, answered first. Open with a 40 to 60 word answer that stands alone if lifted with no context. Then earn the rest of the page. If your first paragraph is scene setting, you have donated the citation to whoever was more direct.

Write in self-contained chunks. Every H2 section should make sense pulled out on its own, which in practice means roughly 130 to 170 words, no "as mentioned above", and the subject named rather than pronouned. Question-shaped headings help, because they match the shape of what the engine is looking for.

Attribute everything numeric. This is the highest-leverage habit in the whole discipline and the KDD study backs it: quotations, statistics and citations were the three best-performing edits of the nine tested. "Conversion improved significantly" is unquotable. "Conversion rose from 2.1% to 3.4% over six weeks (internal data, n=1,840)" is a sentence an engine can safely repeat.

Use tables and lists for anything comparative. A table is pre-chunked, unambiguous data. Comparison content is cited disproportionately across every dataset published on this, which is a strategic hint as much as a formatting one: build the "X vs Y", "best X for Y" and "alternatives to X" pages for your category, and be genuinely fair in them.

Here is the difference in practice, on the same underlying facts:

Where the models actually look

The hardest thing for an SEO team to accept about GEO is how much of it happens on pages you do not control. Ask any engine to recommend software in your category and watch what it cites: roundups, review platforms, Reddit threads, industry media, YouTube. Your own site tends to appear as corroboration, not as the basis of the recommendation. That is rational behaviour. A brand's own claim that it is the best option is the least informative source available.

So the off-site programme is the strategy, not the nice-to-have. Run it in this order. First, identify the gaps by finding out what is already being cited for your category's prompts, because that list is your actual target list and it is usually nothing like your backlink prospect list. Second, get correctly listed in the roundups that appear: being absent from "best X tools" articles the engines lean on is a straightforward, fixable gap, and most of those pages accept submissions or corrections. Third, be described consistently everywhere you appear, since an engine assembling a category from twelve sources rewards the brand those sources agree about. Fourth, earn a presence where people ask each other, which mostly means Reddit and YouTube, by being useful in public rather than by posting about yourself.

Brand mentions matter more than links here, which inverts a decade of SEO instinct. An unlinked paragraph describing what you do, on a page an engine retrieves, contributes more to being named than a followed link from a page it never reads. Monitoring mentions across the engines is how you tell which is which.

A 90-day GEO roadmap: how to implement GEO on a website

This is the sequence we run, and the ordering is deliberate: measurement first, because without a baseline you cannot tell improvement from the engines' day-to-day noise, and off-site last, because it is the slowest to compound and the least useful to start before you know what is being cited.

How to measure the success of GEO campaigns

GEO measurement has one hard problem: your primary outcome happens inside somebody else's product, leaves no referrer, and is non-deterministic. Analytics cannot see it. ChatGPT traffic mostly lands in your reports as direct. So the measurement has to be sampled, the way brand tracking always has been: fix a prompt set, re-run it on a schedule, and read the trend rather than any single answer. Treat it as testing rather than reporting: you are running the same experiment repeatedly and watching the result move.

Four rules keep the numbers honest. Freeze the denominator, because a visibility percentage over a changing prompt set is not a trend. Never score a failed API call as a non-mention, which is the most common way a GEO dashboard invents a decline that did not happen. Re-run on a fixed cadence rather than when someone is curious. And separate mentions from citations in your reporting, because conflating them hides which one you actually improved.

Getting from a percentage to a revenue figure

Nobody funds a mention rate. The bridge that works is a self-reported attribution question on your signup or demo form ("how did you hear about us?" with an AI assistant option), because it is the only instrument that catches the journey analytics loses. Run it alongside the sampled visibility trend and you have a defensible story: visibility rose on these prompts, AI-sourced self-reports rose behind it, those deals closed at this rate. AI search visibility KPIs works through that chain with a full example, and AI visibility metrics and KPIs is the shorter reference for each definition.

One caution on sentiment, since every vendor sells it: scoring whether a mention was positive means asking a model to grade another model's prose, and the result is unstable enough between runs that a change in the score usually tells you nothing about a change in the world. We deliberately do not report it. Read the answer text instead, which takes ten minutes a week and is genuinely informative.

Generative engine optimization software and platforms

The tool category is about two years old and crowded, and whether a product calls itself a GEO checking tool, an optimizer or a visibility platform tells you nothing: they differ less than the pricing does. Every one of them does the same core thing: run a prompt set against one or more engines on a schedule and store what came back. What separates them is whether you can trust the numbers and whether the output tells you what to do.

Four questions cut through a demo faster than a feature matrix, and the first is really about data transparency. Where does the data come from? A vendor that will not say whether it queries a consumer product, a public API or a grounded model is asking you to report a number you cannot explain to your CFO. Can you see the raw answers? If you only get a score, you cannot audit a decline, and you will get declines that are measurement artifacts. What happens on an API failure? The correct behaviour is to record nothing and retry; scoring it as a non-mention manufactures dips. Does it tell you what to change? A dashboard that reports you are at 34% without naming the prompts you lost or the domains that beat you has handed you a number, not a plan.

On free tools: a genuinely free tier that runs a real prompt set is the right way to start, because it gets you a baseline before you have committed budget to a category you are still learning. Our comparison of the AI visibility tools on the market goes through the options including the ones we do not sell, and the engine-by-engine trackers cover the platform-specific setup: ChatGPT, Gemini, Perplexity, Google AI Mode and AI Overviews.

CITED is our answer to those four questions. The free check researches your category's real buyer prompts, runs them through a grounded engine, and returns your visibility score, the competitors named instead of you, the exact domains the answers were built from, and a prioritised action plan. Every underlying answer stays visible, and a failed call is recorded as no data rather than as a miss. The recommended starting point is whichever tool gets you a frozen prompt set fastest. If you would rather a GEO expert ran it as a service, that is our GEO service, and the process behind it is written up in full.

Does GEO work for all types of websites?

It works wherever people ask an assistant for a recommendation before they buy, which is most considered purchases and very little impulse retail. The mechanics do not change by industry. What changes is which prompts matter and which third-party sources the engines trust for your category, and that difference is large enough to make a generic GEO plan mostly useless. An ecommerce site and a B2B platform can run the same process and end up working on entirely different pages, and the gap between what small businesses and what enterprises need is governance rather than technique.

A note on financial services and other regulated markets

Regulated categories are the highest-stakes place to do GEO and the easiest to do badly, because the same content that wins citations can create compliance exposure. Three things are worth getting right before you publish anything.

First, credentials are not decoration here. A named author with a real, verifiable qualification is the difference between content an engine will build a financial answer on and content it will pass over, and it is also what a quality rater is looking for. Second, be careful with authority-claiming structured data. There is no GEO benefit to marking your pages up as regulated financial or medical content, and doing it without the credentials to back it up is a genuine risk; Organization, Article and FAQPage are the right set. Third, precision beats confidence. "Rates from 4.2% APR, representative example as of August 2026" is both quotable and compliant. "Market-leading rates" is neither. The GEO-optimal sentence and the compliant sentence turn out to be the same sentence, which is a happier outcome than most teams expect going in.

Challenges in implementing GEO: five ways programmes fail

The GEO audit checklist and implementation checklist

Twelve checks, in the order worth doing them. If you are starting from nothing, the first four will tell you whether you have a visibility problem or an invisibility problem, and those need different plans.

The future of generative engine optimization

Two things look reasonably safe to bet on. The measurement will get harder before it gets easier: as answers personalise, a single visibility number for a brand becomes less meaningful and sampled, segmented measurement becomes the only defensible approach. And the tactical layer will keep getting commoditised. Answer-first structure and question headings are already table stakes rather than an edge, exactly as meta descriptions were in 2005.

What that leaves is durable and unglamorous: being genuinely worth recommending, being described consistently by people who are not you, and publishing things nobody else can because you have first-hand data. That is not a satisfying prediction for anyone hoping GEO is a settings change. It is also the reason the work compounds instead of expiring at the next model release.

If you want to know where you stand before committing to any of this, start with the baseline: run the free AI visibility check, or read our GEO process for how we run the 90 days above as an engagement.

GEO FAQ