GEO Best Practices for 2026

Search for "GEO best practices" and you'll find the same list a dozen times over: answer the question first, add statistics, use schema, build authority, keep content fresh, get mentioned on other sites. Some of that advice is sound. Some of it is a guess repeated so often it sounds like a fact. Almost none of those lists tell you which is which.

That matters, because your time is limited. If you're going to rework pages for AI answers, you want to start with the changes that have real evidence behind them and treat the rest as reasonable bets, not rules.

So this guide does something the usual lists skip. It grades each common GEO practice by how well it's actually supported, using three kinds of sources: the 2023 study that started the field, the follow-up research published since (pulled together in a critical review released in July 2026), and what Google and Bing now say officially. Where the evidence is strong, we say so. Where it's thin, we say that too, and explain why some thin-evidence practices are still worth doing.

If you need the basic definition of GEO first, and how it relates to AEO and ordinary SEO, start with our AEO vs. GEO comparison. This post assumes you know what the term means and want to know what to actually do.

Where "GEO Best Practices" Come From

Nearly every GEO list traces back to one paper: "GEO: Generative Engine Optimization," by researchers at Princeton and IIT Delhi, first posted in November 2023 and accepted to KDD 2024, a major data-science conference.

The researchers built a test set of 10,000 questions, which they called GEO-bench, and a test answer engine that worked like the real ones: search the web, pull in a handful of pages, write an answer, cite sources. Then they took the pages being cited and rewrote them nine different ways to see which rewrites made a page show up more in the answers. The nine were:

• Authoritative: a more confident, persuasive tone
• Statistics Addition: swapping vague claims for numbers
• Keyword Stuffing: repeating words from the question
• Cite Sources: adding references to credible sources
• Quotation Addition: adding quotes from relevant people or sources
• Easy-to-Understand: simpler language
• Fluency Optimization: smoother, better-flowing text
• Unique Words: adding unusual terms
• Technical Terms: adding specialist vocabulary

The winners were the rewrites that added checkable evidence. Adding quotations, adding statistics and citing sources came out on top, alongside improving fluency. In the paper's own words, the best methods could "boost visibility by up to 40% in generative engine responses." Keyword stuffing, the oldest trick in search marketing, did the opposite: it made pages slightly less visible than doing nothing.

Two details from the paper rarely make it into the lists, and both matter.

The effect depended heavily on where you started. The researchers looked at pages by where they ranked in the search results the engine drew from. For the page ranked first, the three evidence rewrites (sources, quotations, statistics) actually cut its visibility, by roughly 21% to 30%. For the page ranked fifth, the same rewrites roughly doubled it. That makes the methods look less like a universal boost and more like a way for a page that's already in the running to catch up.

The effect changed by topic. Quotations helped most on history and people-focused questions. Statistics helped most on law, government and opinion questions. Citing sources helped most on plain factual questions. The authors wrote that this shows "the need for domain-specific optimization methods." In other words, no single recipe won everywhere.

That's the origin of the standard list. The question is what happened next.

What Three Years of Follow-Up Research Changed

Between late 2023 and mid-2026, GEO grew from one paper into a small research field. In July 2026, Olivier Martinez of Sciences Po published "Optimizing Visibility in Generative Engines," a critical review of that work, covering studies up to July 14, 2026. It's a preprint, meaning it hasn't gone through formal peer review yet, so treat it as a careful summary rather than the final word. But it's the most complete look at the evidence so far, and its main conclusion is blunt.

According to the review, later research "narrowed the scope" of the original findings in four ways:

1. Relevance and position matter most. How closely a page matches the question, and how high it sits among the pages the AI pulls in, are "the most reproducible levers." One study the review covers found that moving a source higher in that list "has a greater effect than most rewrites."
2. General tricks don't carry over well. A rewrite that helps on one engine or one topic often does nothing on another.
3. Competition wears the gains down. When several sites all apply the same rewrites, the advantage shrinks, because being cited is partly a zero-sum game: there are only so many citations per answer.
4. Rewriting for citations can backfire. Changes aimed at getting quoted can make a page less likely to be found in the first place.

The review's own summary of what is safe to recommend is worth quoting in full: "produce a relevant, comprehensive, verifiable, clearly structured, and technically retrievable page; then measure retrieval, citation, and fidelity separately." In plainer words: make a page that answers the question fully, backs its claims, is easy to follow and can actually be reached, then check separately whether AI finds it, cites it and describes it correctly. It adds that this looks more like "high-quality information engineering than keyword manipulation."

That's less exciting than a list of hacks. It's also a much better guide to where to spend your effort.

The Three Gates Every Page Has to Pass

One idea from the review makes the rest of this guide easier to follow. For a page to get cited in an AI answer, three separate things have to go right, in order:

1. The engine has to decide to search at all. Many questions get answered without any web lookup, and some search features only switch on for certain question types.
2. Your page has to be pulled in. The engine searches, gathers a shortlist of pages, and your page has to make that shortlist.
3. Your page has to be cited. Out of the shortlist, the engine picks which sources to actually name.

Most GEO advice, including almost everything the 2023 paper tested, is about gate three: making a page that's already on the shortlist more likely to be cited. Gates one and two get far less attention, even though a page that never makes the shortlist can't be cited no matter how well it's written.

The review's figures show how much gate one alone varies. One study of 55,393 trending searches found Google showed an AI Overview for 13.7% of them overall, but for 64.7% of searches phrased as questions. So the wording of the question decides whether there's an AI answer to appear in at all.

Gate two varies by engine, too. In one audit the review covers, only 26% of the domains cited by Bing Chat and Perplexity were cited by both. Another found that 53% of the domains cited by Google's AI Overviews didn't appear in Google's regular top 10 results. As the review puts it, these results "refute the notion of a global GEO ranking." There is no single AI search leaderboard to climb. Each engine, and each mode within an engine, picks its sources its own way.

Keep those three gates in mind as you read the grades below. A practice that helps at gate three is worth little if you're failing at gate two.

GEO Best Practices, Graded by the Evidence

Strong evidence

1. Match the page to the real question. This is the most consistently supported finding in the whole field. A page that directly and completely answers what people actually ask beats a page that's been cleverly rewritten but only half-answers it. In practice: figure out the real questions your customers ask, make sure each important question has a page that answers it head-on, and don't make the reader (or the AI) dig for the answer. This is also exactly what Google's own guidance says. Its generative AI guide, last updated in July 2026, says that "creating content that people find unique, compelling, and useful will likely influence your website's presence in generative AI search in the long run more than any of the other suggestions in this guide."

2. Make sure AI systems can actually reach and read the page. This is gate two, and it's where many sites fail without knowing it. If an AI crawler is blocked by your robots.txt file, by a firewall setting at your hosting company, or by a page that only fills in its text after JavaScript runs, your page never makes the shortlist. Nothing else on this list helps until this is fixed. Our guide to AI crawler access walks through each check. This one isn't glamorous, but it's the cheapest win available, because a single blocked crawler can undo every other improvement.

Moderate evidence

3. Put checkable facts on the page. The original finding that statistics, quotations and cited sources help has held up reasonably well. The review rates "statistics, definitions, comparisons, prices, dates, and references" as having a plausible advantage, and notes one large controlled experiment that found real effects for explicit prices and recent dates. The key word is checkable. A real number from your own data, a clear price, a named source with a link: these give an AI something specific to repeat and attribute. An invented or vague statistic does the opposite, and risks the kind of wrong answer covered below.

The review adds a useful limit: the effect depends on what the person wants. A recent date helps with a time-sensitive question, but not necessarily with a stable definition. A quotation may help on a history question but add nothing to a pricing question. So add evidence where it answers the question better, not everywhere by default.

4. Keep time-sensitive facts current. Following from the point above, for anything that changes (prices, opening hours, product details, "best of" comparisons, anything with a year in it), an up-to-date page with a visible date has support behind it. For evergreen explanations, freshness matters much less. Refresh the pages where being out of date would make the answer wrong, and don't bother re-dating pages just to look new.

5. Write clearly for people first. Fluency and plain language did well in the original study, and the review's advice is simply to "optimize for the user first." Clear writing also protects you at gate three for a reason that has nothing to do with rankings: an AI that can easily tell what your page says is less likely to misquote it.

Unproven, but still worth doing

6. Make your identity impossible to mistake. This is where schema markup and a complete entity profile (the structured facts about who your business is, what it does, and where it's listed) come in, and it deserves an honest note. No study in the review isolates structured data as a cause of more citations. The review itself lists structured data among the factors future research still needs to separate out. A controlled study by Ahrefs in May 2026, covered in our look at schema in 2026, found no meaningful citation lift after pages added schema. And Google says plainly that structured data "isn't required for generative AI search, and there's no special schema.org markup you need to add," while still recommending it as part of normal search work.

So why is it on this list at all? Because citation counts aren't the only thing at stake. The review cites an early study of four answer engines where only 51.5% of the sentences in AI answers were fully supported by the sources they cited, and only 74.5% of citations actually backed up the sentence they were attached to. AI answers get things wrong, including things about businesses. When an AI answers a question about you by name, clear, consistent facts about your business, stated the same way on your site, in your schema and on your listings, give it something solid to repeat instead of a guess. That's a job about accuracy, not ranking, and it's the reason we treat an entity profile as core work rather than an extra.

7. Earn mentions on other sites. Many lists put "get mentioned in the press, in reviews, in round-ups" near the top. The review is cautious: one study found that AI answers lean heavily on earned media (news and independent sites rather than brand-owned ones), but that study was observational and industry-funded, so it shows a pattern, not a cause. The review's reading is that third-party mentions "may expand the ecosystem of retrievable evidence," meaning more places for an engine to find you. That's a sensible reason to pursue them, as long as the mentions are genuine.

Weak, or actively harmful

8. A more "authoritative" tone. Rated "weak and unstable" in the review, with a warning that it "may conflict with credibility." Sounding confident is not the same as giving evidence, and an AI that's comparing sources can tell the difference.

9. Formatting recipes on their own. Breaking everything into tiny chunks, forcing a fixed heading pattern, or adding summary boxes to every page shows "poor generalization" in the review (it rarely works outside the test where it was found), with only occasional local gains. Google's guide agrees: "There's no requirement to break your content into tiny pieces for AI to better understand it," and "you don't need to write in a specific way just for generative AI search." Good structure helps readers. It isn't a special AI key.

10. Keyword stuffing. Null or negative in every test that's looked at it, including the original paper, where it made pages less visible.

11. Special AI files for Google. Google's guide says you don't need "new machine readable files, AI text files, markup, or Markdown" to appear in its AI features, because Google Search doesn't use them. Other engines may treat files like llms.txt differently, and our llms.txt comparison covers where it may still help. Just don't expect it to move anything in Google.

12. Trying to trick the AI. Hidden instructions aimed at language models, fake reviews, and pages written to manipulate recommendations are a real category now. The review covers research showing these attacks can work in lab settings. Search engines have noticed: in February 2026, Bing updated its guidelines to name GEO directly and added a full section on prompt injection and AI manipulation. Whatever short-term lift these tricks buy, they put your whole site at risk.

Being Recognized Is Not the Same as Being Recommended

One finding in the review changes how you should think about all of the above. A study of 112 startups, summarized in the review, tested two kinds of question. When asked about a product by name, ChatGPT recognized 99.4% of them. When asked a general question where that product would be a fair answer, ChatGPT brought it up only 3.32% of the time. Perplexity dropped from 94.3% to 8.29%. The review is careful to note this is a single preprint using two models, but the gap is enormous.

In plain terms: AI knowing who you are when someone types your name is one problem. AI suggesting you when someone asks "what's a good option for X" is a completely different, much harder problem.

That split lines up with the practices above:

• The named question ("What does [your business] do? Where is it? How much does it cost?") is mostly about accuracy. Clear identity, consistent facts and a well-built entity profile are what help here, practice 6 above.
• The category question ("Who's a good [service] in [city]?" or "What's the best tool for Y?") is about relevance, retrieval and a wider network of sources: practices 1, 2, 3 and 7.

Most businesses need both, but they're worth checking separately. If AI gets your facts right when asked by name but never mentions you otherwise, your identity work is done and your effort should go into relevance and outside mentions. If it gets your basic facts wrong, fix that first, because being recommended with the wrong phone number or price is worse than not being mentioned.

Measure Each Gate Separately

The review's last piece of advice is to "measure retrieval, citation, and fidelity separately," meaning whether AI finds you, cites you and gets you right, and it's the part most GEO programs skip. A single "AI visibility score" hides which gate is actually failing. Here's how to check each one yourself:

• Can AI reach your pages? Check your server logs for visits from AI crawlers, and check robots.txt and any firewall rules. If the crawlers aren't arriving, stop here and fix that.
• Are you being shown and cited? Google Search Console now has a generative AI performance report, and Bing Webmaster Tools has an AI Performance report (still a public preview). Both are free and come straight from the engines, which makes them more trustworthy than third-party estimates.
• Is what AI says about you correct? Ask the major AI tools the named questions your customers ask about you, and write down what comes back. Wrong prices, old addresses, or mixed-up competitors all point to an identity problem, not a ranking one.

One more caution from the review: evidence on traffic and sales is the weakest part of the whole field. In one study it describes, a website's ChatGPT referrals grew 5.7 times after some pages were optimized, but the pages that weren't touched had grown 3.5 times anyway, because ChatGPT itself was growing. The review's conclusion is that "claims about GEO return on investment clearly outstrip the academic evidence." If a vendor promises a specific lift in citations or revenue, ask how they separated their work from the general growth of AI search.

Common Mistakes

Starting at gate three. Rewriting pages with quotes and statistics while an AI crawler is blocked, or while key pages depend on JavaScript to show their text, is wasted effort. Check access first.

Treating the "up to 40%" figure as a promise. It's in the original paper's summary, but it came from the authors' own test engine, measured on pages that were already in the answer's source list, and later research found such gains often shrink or vanish in other settings.

Applying one recipe everywhere. Even the original paper found different methods worked for different topics. A pricing page, a how-to guide and an "about us" page need different kinds of evidence.

Adding evidence you can't back up. The point of statistics and sources is that they're checkable. A made-up figure can end up repeated by an AI under your name, which is a worse outcome than having no figure at all.

Measuring only one number. If you only track how often you're cited, you can't tell whether a drop came from blocked crawlers, a change in how the engine answers, or a competitor's new page.

Expecting schema alone to get you cited. It won't, and the evidence says so. Its job is to make your identity clear and your facts consistent, which matters for accuracy, not as a citation shortcut.

How to Check Where Your Site Stands

The strongest GEO practices on this list, being reachable, being readable, being clearly identified and giving answers backed by your own facts, are all things you can check on your own site today, because they're under your control. How often an engine chooses to cite you isn't. That split between what you can fix (readiness) and what you can only influence (visibility) is the thread running through readiness vs. visibility, and it's the most useful way to plan GEO work honestly.

A reasonable order of work:

1. Fix access. Make sure AI crawlers can reach and read your important pages.
2. Fix identity. Get your business facts consistent across your site, your schema and your listings.
3. Fix the answers. Make sure each key customer question has a page that answers it directly, with checkable facts where they help.
4. Then earn mentions and measure. Build outside mentions and track each gate separately over time.

The first three steps are exactly what an AI readiness audit checks. Run a free readiness check to see whether AI systems can reach your site, read it, tell who runs it and answer questions about it, with a specific list of what to fix first.