How to Rank in AI Search: A Practical Guide to GEO and AEO

What "ranking" means when there is no results page

The phrase is borrowed and slightly wrong. In an AI answer there is no position one; there is an answer, and sometimes a handful of citations attached to it. Your goal is to be one of those citations — or, where no citations appear, to be one of the sources the answer was synthesised from.

That is a meaningfully different target from a blue link. A search engine ranks pages by how well they match a query and how authoritative they seem. An assistant does something extra: it has to lift a specific claim off your page, be reasonably confident that claim is correct, and be willing to attach your name to it. A page can rank well and still never be quoted, because nothing on it is quotable.

This is why the discipline picked up new names — generative engine optimisation (GEO) and answer engine optimisation (AEO). The names are marketing, but the underlying shift is real: optimising for extraction rather than for clicks.

How AI answers are actually assembled

Understanding the pipeline tells you where you can intervene. Broadly, three mechanisms produce the answers people see, and they overlap.

The first is retrieval-augmented: the assistant issues one or more searches, fetches a set of pages, reads them, and synthesises an answer from what it read, usually with citations. Perplexity, ChatGPT with browsing, and Google's AI Overviews all work broadly this way. This is where your page content matters most directly and most immediately.

The second is parametric: the model answers from what it absorbed during training, with no live retrieval. Here you cannot influence a specific answer, but the long-run presence of your brand, your terminology and your claims across the web is what shaped what it absorbed.

The third is a hybrid where the model has a strong prior from training and uses retrieval to confirm or update it. In practice most consumer AI search is this — which means both the immediate extractability of your page and your longer-term footprint across the web are doing work.

• Retrieval-augmented: your page must be fetchable, parseable and quotable. Wins are fast, sometimes within days of publishing.
• Parametric: your presence across many sources over time shapes what the model believes. Wins are slow and compound.
• Hybrid: consistency between the two matters. If your page contradicts what the model already believes, it may prefer its prior or hedge — either way you do not get cited cleanly.

The single highest-return change: a self-contained answer near the top

If you do one thing, do this. Directly under the heading that states the question, write a two-to-four sentence paragraph that answers it completely, without pronouns pointing backwards, without "as we discussed above", and without depending on any surrounding context.

The reason this works is mechanical. An assistant assembling an answer is looking for a span of text it can lift. A paragraph that stands alone is liftable. A paragraph that begins "This is where it gets interesting" is not, no matter how good the rest is, because pulled out of the page it means nothing.

Write it as if it will be read aloud with no other context, because increasingly it will be. Repeat the subject noun instead of using "it". Include the qualifying condition in the same sentence rather than in the next one. If the honest answer is "it depends", say what it depends on inside the paragraph rather than promising to explain later.

• Weak: "There are a few factors here, and the answer changes depending on your setup — we will get into them below."
• Strong: "A standard espresso shot uses 18 to 20 grams of ground coffee for a double basket and yields 36 to 40 grams of liquid in 25 to 30 seconds. Single baskets use 7 to 9 grams. Adjust grind size first when the shot runs outside that window; adjust dose only after grind is dialled in."
• The strong version can be quoted verbatim by an assistant and remains correct and attributable. The weak one cannot be quoted at all.

Structure the page for a parser, not just a reader

Assistants read HTML structure. Headings tell them what a section is about, lists tell them a set of discrete items exists, tables tell them there are dimensions and values. Prose tells them there is prose.

This does not mean writing badly. It means that where information is genuinely list-shaped or table-shaped, present it that way rather than dissolving it into paragraphs for stylistic reasons. A comparison written as three paragraphs of flowing prose is a comparison an assistant will probably get wrong; the same comparison as a table with labelled rows is one it will get right.

• Use one H1 and a logical H2/H3 hierarchy. Phrase headings as the questions people ask, not as clever labels — "How much does it cost?" beats "Investment".
• Put specifications, steps and comparisons in lists and tables. Semantic HTML, not visually-styled divs.
• Keep paragraphs short and single-idea. A paragraph containing three claims is harder to extract one claim from.
• Front-load every section: the claim first, the elaboration after. Inverted pyramid, per section, not just per article.
• Avoid burying key facts inside images, PDFs or JavaScript-rendered components. If it does not exist in the served HTML, assume it does not exist.

Schema markup: the part most sites skip

Structured data does not guarantee a citation, but it removes ambiguity about what your page contains, and ambiguity is what makes a machine skip you in favour of a page it understood.

Four types carry most of the weight for AI search. Implement them correctly with JSON-LD in the page head, validate them, and make sure the marked-up content is genuinely visible on the page — marking up content users cannot see is a policy violation with real consequences.

• FAQPage — for genuine question-and-answer content. Each question a real question, each answer self-contained. The highest-value markup for AEO because it maps exactly onto the shape of a query.
• HowTo — for step sequences, with each step as a discrete item. Tools, materials and total time as properties.
• Article — with author, datePublished, dateModified, and a real publisher. The date fields matter more than people assume; assistants asked for current information visibly prefer content with fresh, honest timestamps.
• Product — price, availability, specifications and reviews as structured properties rather than as prose. For anything transactional this is the difference between being described accurately and being described vaguely.
• BreadcrumbList and Organization — context about where a page sits and who published it. Cheap to add, and part of how a model decides whether you are a real entity.

Publish what only you can publish

An assistant synthesising an answer from ten pages that all say the same thing has no reason to cite any particular one. It will cite whichever it considers most authoritative, which is usually not you.

The way to be cited is to be the only source of something. Original data from your own operations. A methodology someone else would have to repeat. A number you measured. A limitation nobody else admits. A definition specific enough to be worth quoting.

This is also the only defensible strategy long term. Everything in this article that is mechanical — schema, headings, self-contained paragraphs — will be done by everyone within a couple of years, because it is easy. What will still differentiate you is having something to say that the model cannot get from the other nine pages.

• Original measurements from your own product, tests or customer base. Say how you measured.
• Named methods and frameworks. A model looking for a way to structure an answer will reach for a named thing over an unnamed one.
• Explicit limitations and failure cases. "This does not work when X" is one of the most quotable sentence forms there is, and almost nobody publishes it.
• Honest comparisons that include where you lose. Assistants asked to compare will find the discrepancy if you only publish flattering claims, and hedge or skip you as a result.
• Concrete numbers with units and conditions attached. Vague quantities do not survive extraction.

Be findable: the technical minimum

None of the content work matters if the crawlers cannot reach the page or cannot render it. This is unglamorous and it is where a surprising number of sites silently fail.

Server-side render anything you want cited. Many crawlers execute limited or no JavaScript, and content that only appears after hydration is invisible to them. Check by fetching your own page without JavaScript and reading what comes back.

• Serve content in the initial HTML. If your framework can pre-render or server-render, do it for content pages.
• Check your robots.txt against the AI crawler user agents you care about. Several are separate from the classic search bots, and blocking them is a decision, not an accident — make it deliberately.
• Keep the page fast and the HTML clean. A crawler with a budget gives up on bloated pages.
• Use canonical URLs consistently, and avoid publishing the same content at multiple addresses.
• Maintain a real sitemap and keep lastmod dates honest.
• Consider an llms.txt file at your root — a plain-text index of your most useful pages with short descriptions. It is a young convention rather than a standard, costs almost nothing, and makes your best content easy to enumerate.

Freshness is a ranking factor in a way it never quite was

Assistants are unusually date-sensitive, partly because users ask date-sensitive questions and partly because models are trained to be cautious about stale information. A page dated three years ago is treated as a weaker source than an equivalent page dated last month, even when the underlying facts have not changed.

The response is not to churn out fake updates. Changing a date without changing content is transparent to anyone checking and does nothing for a model comparing the text it retrieved to what it already believed.

The response is a genuine review cycle: revisit your important pages on a schedule, update what has actually changed, remove what is no longer true, add what you have learned, and then update the modified date honestly.

• Put visible dates on content pages, and keep dateModified in your Article schema accurate.
• Review high-value pages quarterly. Fix stale claims, prices, version numbers and screenshots.
• When you update substantially, say what changed. A short "updated: X now works differently" note is itself quotable.
• Remove or clearly mark content that is genuinely obsolete rather than leaving it to be retrieved and cited wrongly.

Consistency across the web: the entity problem

Models build a picture of who you are from everything they have seen, not from your site alone. If your homepage, your directory listings, your social profiles and third-party write-ups describe you differently, the picture is blurry — and a blurry entity gets described vaguely or not at all.

The fix is boring and effective: describe yourself the same way everywhere. Same name, same one-line description, same category, same core claims. This is old-fashioned entity SEO, and it turns out to matter more for AI answers than it ever did for blue links, because a model synthesising a description of you is averaging every source it has.

• Write one canonical description of what you do, in one sentence, and use it verbatim everywhere.
• Keep names, spellings and capitalisation identical across sources. "PromptVibe", "Prompt Vibe" and "promptvibe" can read as three things.
• Get described accurately in places that get crawled: relevant directories, comparison sites, communities where your category is discussed.
• Correct wrong third-party descriptions when you find them. A stale review saying you lack a feature you shipped a year ago is actively feeding the model bad information.
• Maintain Organization schema with sameAs links to your official profiles, which is the most direct way to tell a machine that these accounts are the same entity as this site.

Write for the question, not the keyword

Queries typed into a search box and questions asked of an assistant have different shapes. Search queries are compressed — "espresso dose grams". Assistant questions are conversational and often contain the context, the constraint and the goal — "how many grams of coffee should I use for a double shot if my espresso is running sour".

That has two consequences. First, the long, specific, conversational question is now a viable target rather than something too low-volume to bother with. Second, the constraint embedded in the question — "if my espresso is running sour" — is where the answer differentiates, and pages that only answer the generic version get passed over for the page that addressed the condition.

• Collect real questions: your support inbox, sales calls, community threads, the "people also ask" boxes, and the follow-ups your own users type.
• Give each substantial question its own H2 and its own self-contained answer paragraph, phrased as a person would ask it.
• Cover the conditional branches explicitly — "if you are on a Mac", "if your data is already in a spreadsheet", "if you only have the free tier". These conditions are what assistants match against.
• Include the negative case. "When you should not do this" is retrieved for a whole class of queries nobody else is targeting.

What gets you skipped

It is worth being explicit about the failure modes, because most pages that do not get cited are not missing a technique — they are doing one of these.

The common thread is that each one makes a page either hard to extract from or risky to attribute to. A model that is unsure will hedge, paraphrase without citing, or use a different source. You never see it happen.

• The answer is buried after eight hundred words of preamble about why the topic matters.
• Every paragraph depends on the previous one, so nothing can be lifted.
• The page makes claims with no specifics — "significantly faster", "most users", "industry-leading" — which are unquotable because they say nothing.
• Statistics are cited with no source, or with a source that does not contain them. This is increasingly checkable and increasingly checked.
• The content contradicts what the model already believes and offers no evidence to shift it.
• Key content is rendered client-side and invisible to the crawler.
• The page is transparently promotional, so anything it says about the category is discounted.
• The same content exists on fifty other sites, so there is no reason to pick yours.

Measuring something you cannot see in a rank tracker

AI search visibility is genuinely harder to measure than classic rankings, and the honest position is that nobody has clean numbers. Referral traffic from assistants is under-reported, some interfaces send no referrer at all, and an answer that satisfies the user without a click leaves no trace in your analytics.

What you can do is measure directly and repeatedly rather than waiting for a dashboard to tell you.

• Keep a list of twenty to thirty questions where you should be a legitimate source. Ask them across the major assistants on a fixed schedule and record whether you appear, and whether what they say about you is correct.
• Watch for referral traffic from assistant domains in your analytics, and accept it as a floor rather than a true count.
• Track brand-name queries in your search console. Rising branded search alongside flat organic clicks is a signal that people are discovering you inside an assistant and then looking you up.
• Log the wrong answers. An assistant confidently stating something false about you is a content gap you can fix, and it is the single most actionable finding this exercise produces.
• Do not buy a "GEO score" that cannot explain its methodology. The measurement problem is real and a number that hides it is worse than no number.

How this relates to normal SEO

The most common question, and the answer is undramatic: classic SEO gets you into the candidate set, and extractability decides what happens next. Retrieval-augmented assistants mostly find pages through search infrastructure, so a page that does not rank at all rarely gets retrieved at all.

So none of the fundamentals go away. Crawlability, internal linking, page speed, genuinely useful content, and links from places that matter still determine whether you are in the pool. What changes is that being in the pool is no longer the finish line.

The practical implication is reassuring: this is not a second strategy running alongside your existing one. It is a set of editorial and technical habits layered onto the content you were already producing — write the answer first, structure the page, mark it up, be specific, keep it current.

A checklist you can apply to a single page today

Pick one page that should be cited and is not, and work through this. It takes about half an hour and is the fastest way to understand what the work actually involves.

• Is there a self-contained answer within the first screen, in two to four sentences, that would still be correct quoted alone? If not, write one.
• Is the H1 the question a person would ask? Are the H2s the follow-up questions?
• Can each section's first sentence be extracted and stand alone?
• Is anything list-shaped or table-shaped currently trapped in prose? Move it.
• Does the page contain at least one specific, sourced fact that no competitor page has?
• Are there vague claims that should be numbers, or numbers with no source that should be removed?
• Is there FAQPage or HowTo schema where it genuinely applies, and does it validate?
• Are datePublished and dateModified present and honest?
• Does the served HTML contain the content when JavaScript is disabled?
• Is there an explicit statement of when this does not apply, or who this is not for?
• Ask three assistants the question this page answers. Are you cited? Is what they say about it correct?

The uncomfortable part

A strategy that succeeds at getting cited will sometimes mean the user gets their answer without visiting you. That is a real cost and it is worth being clear-eyed about it rather than pretending the traffic maths is unchanged.

The businesses handling this well are the ones where the citation itself has value: the answer names them, positions them as the authority on the thing they sell, and the click that matters comes later, when the reader has a problem the answer cannot solve on its own. That is a brand and product outcome, not a pageview outcome.

It also changes what is worth writing. Content whose entire value was ranking for a question and showing ads next to the answer is in structural decline. Content that demonstrates a capability, documents something only you know, or leads to a thing you actually do is not. If your content strategy was built on the first model, the honest response is to change the strategy rather than to optimise harder.

What to do first

Sequenced by return on effort, for a site that has done none of this.

Start with the pages that already get search traffic. They are already in the candidate set, which is the expensive part; making them extractable is the cheap part, and the improvement shows up fastest.

• 1. Add a self-contained answer paragraph to your top ten pages. Highest return, lowest effort, do it this week.
• 2. Rewrite headings on those pages as questions. An hour of work.
• 3. Add FAQPage schema wherever you have real Q&A content, and validate it.
• 4. Verify your content is in the served HTML without JavaScript, and fix it if not.
• 5. Audit for vague claims and unsourced statistics. Replace with specifics or cut.
• 6. Write your one canonical self-description and propagate it everywhere.
• 7. Set the twenty-question baseline and re-run it monthly.
• 8. Then, and only then, start producing new content built this way from the start.

Related articles