Most GEO advice is a list of tactics: add FAQ schema, publish an llms.txt file, get on Reddit, write listicles. Some of it works. Some of it has never been shown to work. Almost none of it tells you where to start, how to know whether it is working, or what to do when one AI model responds and the others do not.
This guide takes a different approach. It treats Generative Engine Optimization the way a good analyst would: measurement first, then the changes most likely to move the measurement, then measurement again. Seven steps, in order, with the evidence behind each. If you follow them, you will not just be doing GEO. You will know whether it is working.
Generative Engine Optimization (GEO) — The practice of increasing the likelihood that AI answer engines recommend, cite, and accurately describe a brand when users ask questions in its category. It differs from SEO in that the target is not a ranking position but inclusion in a synthesised answer, and the evaluators are several different AI systems rather than one search algorithm. For the full comparison, see GEO vs SEO.
Step 1: Define the Prompts That Matter
Every GEO program starts with a list of questions, and the quality of that list determines everything that follows. The mistake most teams make is starting with their brand name. "What is Acme?" is a question almost nobody asks an AI. "What is the best expense management platform for a 500-person company with international travel?" is a question buyers ask every day, and it is the one where your brand is either on the shortlist or absent.
Build the list in three tiers:
- Problem-statement prompts. The buyer describes a need without naming a category or vendor. These are the highest-value prompts and the hardest to win.
- Category prompts. "Best X for Y," "top X vendors," "X vs Y." Listicle-style questions where AI produces a shortlist.
- Brand prompts. Questions that name you or a competitor. These matter for accuracy and sentiment more than for discovery.
Twenty to fifty prompts is enough to start. Tie each one to a stage of the buying journey, and mark the ones that correspond to revenue.
Step 2: Baseline Across Multiple Models
Run every prompt through several AI models, not one. Our own data on this point is blunt: across more than a hundred categories, five leading models named the same market leader only 38% of the time, and a typical category produced ten distinct vendors across the five models' top-five lists. A baseline from one model is a baseline of one opinion.
For each prompt and model, record which vendors are named, in what order, and how each is described. Then repeat the exercise, because AI answers are probabilistic. The Columbia Journalism Review's Tow Center noted in its 2025 study of AI search engines that identical prompts frequently produce different outputs on re-run; a single pass is a sample, not a measurement.
This is the step most teams skip, and it is the step that makes every later decision possible. If you cannot do it manually at the scale you need, this is what automated multi-model measurement exists for.
Step 3: Fix Entity Consistency
Before creating anything new, fix what AI already believes about you. The Columbia Journalism Review's Tow Center found in 2025 that leading AI search tools answered more than 60% of its test queries incorrectly and rarely acknowledged uncertainty when they did. In our own reviews of AI answers across categories, the most common errors are stale pricing, product names that have been merged or split, and corporate facts that were true two years ago. Every one of them originates in a source the model read.
Consistency means your company name, product names, category description, headquarters, founding facts, and pricing are identical across your website, your structured data, LinkedIn, Crunchbase, review sites, directories, and press. Where they differ, AI models resolve the conflict unpredictably, and different models resolve it differently. This is unglamorous work, and it is the highest-leverage step in the program.
AI cannot recommend a brand it cannot describe. Before you ask to be cited, make sure there is one consistent thing to cite.
Step 4: Create Citable Content
Now create content, and create it to be extracted rather than skimmed. The only peer-reviewed study of GEO tactics, from researchers at Princeton and Georgia Tech published at KDD 2024, tested nine content strategies across ten thousand queries. Three stood out. Adding statistics, adding quotations from credible sources, and adding citations to external sources each improved visibility in generative engine answers by roughly 30 to 40 percent. Keyword stuffing made things worse.
The same study found that these tactics helped lower-ranked content far more than top-ranked content. GEO is a catch-up strategy as much as a leadership strategy, which is good news for challengers.
In practice, citable content means:
- Answer-first structure. The direct answer to the prompt in the first paragraph, then the supporting detail.
- Clear definitions. A bolded term followed by a one-paragraph definition, the format AI engines lift most readily.
- Original numbers. Your own data, survey, or benchmark, stated plainly with dates and methodology.
- Comparison pages. Honest "X vs Y" and "best X for Y" content, because commercial prompts are disproportionately answered from list and comparison formats.
- Depth on the topic, not the brand. AI engines assess topical authority; a site that answers every question in its category is cited more than one that only describes its product.
Step 5: Strengthen Third-Party Corroboration
AI engines weight sources differently, and none of them rely only on your site. Semrush's three-month study of more than 230,000 prompts found Reddit, Wikipedia, LinkedIn, YouTube, and Forbes to be the most-cited domains across ChatGPT, Google's AI Mode, and Perplexity, with sharp swings by platform and by month. Profound's analysis of 680 million citations found Reddit accounted for under 2% of ChatGPT citations but nearly 7% of Perplexity's. A 5W Research audit published in May 2026 found Wikipedia and Reddit together drive more than a quarter of ChatGPT's citations in the United States.
The implication is that the same claim needs to exist in several places. Reviews on the platforms your category uses, coverage in the trade press models cite, comparison content on third-party sites, and a consistent presence on LinkedIn, which is among the most-cited domains for professional queries. Look at your Step 2 baseline to see which sources each model actually cites in your category, and prioritise those.
Step 6: Keep Content Fresh and Crawlable
Freshness matters more to AI engines than to traditional search. Ahrefs' analysis of nearly seventeen million cited URLs found AI assistants cite content roughly 26% newer than Google's organic results, with ChatGPT showing the strongest preference for recent pages. A quarterly refresh cadence for your priority pages is a defensible planning assumption.
Crawlability is the other half. Confirm that your robots.txt allows the crawlers you want to be cited by. OpenAI, Anthropic, Perplexity, and Google each operate separate bots for training, search, and user-triggered fetches, with different behaviours. Blocking a training crawler to protect your content can also block the search crawler that would have cited you.
One caution: do not confuse activity with effect. An llms.txt file is cheap to publish, but no major AI provider has confirmed that it reads them, and there is no published evidence that they influence citations. Publish one if you like. Do not count it as a strategy.
Step 7: Measure, Compare, Repeat
Re-run the Step 2 baseline on a fixed schedule, monthly at minimum, and track three things per prompt and per model: whether you are named, where you rank, and how you are described. Compare against named competitors, not against your own past score alone.
Then read the divergence between models. If the retrieval-based engines move first and the training-data models do not, your new content is working and the rest will follow. If one model's view of your brand diverges sharply from the others, you have an entity problem in a source that model favours. If every model moves together, the market narrative itself has shifted. Each pattern calls for a different response, and only multi-model measurement can tell them apart.
A GEO strategy is a loop, not a launch. Prompts define the target, a multi-model baseline shows where you stand, entity consistency and citable content move the needle, third-party corroboration and freshness sustain it, and measurement tells you what to do next. Teams that run the loop quarterly consistently outperform teams that run a one-time optimisation project.
What to Expect
Retrieval-based engines respond within weeks. Training-data models respond within months, and only after the change has propagated across the sources they learn from. Categories where AI models already agree on a leader are hard to break into; categories where they disagree are open, and in those the first brand to give AI a consistent, specific story tends to become the default answer. Know which kind of category you are in before you set expectations, and you can see that for any market on the QuadrantX Explore page.