When someone asks ChatGPT, Perplexity, or Google a question, the answer increasingly arrives as a written paragraph rather than a list of links. Inside that paragraph, a few sources get named or linked. Getting cited means being one of them. It is not a ranking you can buy, a tag you can add, or a trick you can run once. It is the result of three things lining up: the engines can reach your page, your page answers the question better than the alternatives, and the engines trust what you say enough to repeat it with your name attached.
This guide explains how that works, engine by engine, using only what the platforms themselves have published. OpenAI, Perplexity, and Google have all documented the mechanics in their own help pages and developer guides, and those documents agree on more than the marketing around this topic suggests. Where they disagree, or where they say nothing at all, this guide says so. The honest headline is this: no one can promise you a citation. OpenAI says placement is not guaranteed. Google says its AI features are rooted in its ordinary search ranking, so the work that earns a citation is mostly the work that already earns a good ranking, done with unusual clarity. What follows is the practical version of that: what to open, what to write, what to stop believing, and how to tell whether any of it is working.
Checked 4 November 2026. The platforms change these systems often, so treat the crawler names and report names here as current at that date and re-check the linked documentation before you act on a detail that matters to your site.
1. The short answer: what getting cited actually means
A citation in an AI answer is not a prize for having the best website. It is evidence the engine used while writing. When ChatGPT answers with search, it may attach citations and a Sources panel. When Perplexity answers, it attaches numbered citations to the sources behind each claim. When Google shows an AI Overview, it shows supporting links to the pages behind the summary. In every case the same pattern holds: the engine read something, wrote something based on it, and pointed at where it came from.
That is a different contest from classic search, and the difference matters for how you judge success. A ranking is stable: your page is third for a query, today, in this country, and anyone can check it. A citation is probabilistic: the same question asked twice can produce different answers, cite different pages, or cite nobody at all. The engine generates a fresh answer each time, and its choice of sources can shift with the wording of the question, the user’s location, and the day. So the goal is not to win one answer. It is to raise the odds across many answers, over time, for the questions your audience actually asks.
It also helps to separate three outcomes that people often blur together. Being indexed means an engine has your page in its collection. Being retrieved means the engine pulled your page up as a candidate for a particular question. Being cited means it actually used your page in the answer and pointed at it. Each step has its own failure mode, and the fixes are different. A page that is not indexed needs technical work. A page that is indexed but never retrieved needs better matching to real questions. A page that is retrieved but never cited needs clearer, more quotable, more trustworthy content. Most guides skip this distinction and promise the third outcome while only fixing the first. This guide works through all three in order.
One more distinction is worth making early, because it shapes everything else. There are two kinds of visibility inside an answer: being named and being linked. A brand can be mentioned in the text without a link, and a page can be linked without the brand being discussed. Both have value, but they are not the same, and you should decide which one you are chasing for each page. A product page wants the link. A brand building reputation wants the mention to be accurate and favorable. If you have not read our companion piece, What Is GEO? Generative Engine Optimisation, Explained, it sets out this shift from rankings to mentions in more detail.
<img src="
” alt=”A single glowing answer card rising above a row of plain link cards on a desk, illustrating one citation standing out from a list of links”>
2. How AI answers are built: retrieval first, writing second
Almost every AI answer system works in two stages, and understanding the split explains most of the advice in this guide. First, the system retrieves: it turns the user’s question into one or more searches, fetches candidate pages, and picks the passages most relevant to the question. Second, it writes: a language model composes an answer from those passages, plus its own trained knowledge, and attaches citations to the sources it leaned on. Google gives this technique a name in its own guide: retrieval-augmented generation, or RAG, also called grounding. The model does not simply recite what it memorized in training. It reads fresh pages and writes from them, which is why a new page can be cited days after publication and why an outdated page can be quoted with total confidence.
Google also names a second mechanism that matters for writers: query fan-out. When you ask one question, the system may silently issue a set of concurrent, related queries to gather more information. Google’s own example is a lawn full of weeds: the fan-out queries might cover the best herbicides for lawns, removing weeds without chemicals, and preventing weeds in lawns. OpenAI describes a similar process for ChatGPT search: it typically rewrites your question into one or more targeted queries, sends them to search providers, and may send follow-up queries after reading the first results. The practical consequence is that your page does not need to match the user’s exact words. It needs to answer one of the sub-questions the system actually asks, clearly enough that the passage survives on its own.
This two-stage design explains why two common fears are both wrong. The first fear is that the model already “knows” everything and your page is irrelevant. It is not: for current information, specific facts, and niche expertise, the retrieval stage is doing real work, and the pages it fetches genuinely shape the answer. The second fear is that the model invents everything. It can, and OpenAI says so openly, warning that search results and citations can be incomplete, outdated, or incorrect. But invention is the failure mode, not the design. The design is reading, then writing. Your job is to make the reading easy and the writing safe: passages that are accurate, complete in themselves, and hard to distort when lifted out of context.
There is one more consequence worth noting. Because the answer is assembled from passages, the unit that gets cited is rarely your whole page. It is a paragraph, a list, a table, or a definition. A page can be excellent overall and still never be cited if its best material is buried in the middle of a long argument, split across sections, or wrapped in language that only makes sense after reading three preceding paragraphs. The rest of this guide is, in large part, about making your best passages citable on their own. For the broader question of whether this work replaces classic SEO or sits on top of it, see AEO Vs SEO: How Answer Engines Change The Game For Brands.
3. ChatGPT: how search and citations work
ChatGPT’s search is the system most people mean when they talk about getting cited by AI, so it is worth describing exactly what OpenAI says it does. According to OpenAI’s help page on searching the web with ChatGPT, the system may search automatically when a question would benefit from current information, and the user can also start a search manually from the tools menu. Web search is available on every ChatGPT plan, from Free upward, and even to people who are not signed in. That breadth matters: your cited page can be read by an answer shown to a signed-out user, not only to subscribers.
The process OpenAI describes has four visible steps. First, ChatGPT decides whether the question needs fresh information at all. A question about a timeless concept may be answered from training alone, with no search and no citations. A question about a recent event, a current price, or a specific product almost certainly triggers a search. Second, it rewrites the question into one or more targeted queries and sends them to search providers it partners with, sometimes following up with more specific queries after reading the first results. OpenAI’s own example is a biotech researcher asking about drugs targeting CCR8 for cancer: the system might first query “CCR8 immunotherapy drug development 2025” and then follow up with something narrower like “CHS-114 conference 2025.” Third, it reads the results and writes an answer. Fourth, it attaches citations: responses that use web search may include inline citations, and a Sources button, when available, lists cited sources and other relevant links. Images in a response link back to their sources too.
Two honest warnings come from OpenAI itself. The first is that “ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed.” OpenAI does not publish those factors, and it does not publish what share of cited pages comes from its own crawl versus its search partners. Anyone selling you a precise ChatGPT ranking formula is selling something OpenAI has not disclosed. The second warning is that “search results and citations can be incomplete, outdated, or incorrect,” and OpenAI’s advice is to open the cited source, check that it supports the answer, and check when it was published or updated. That advice cuts both ways for publishers: a citation is only as good as the page behind it, and a page that is out of date or misleading will eventually be caught by the very readers the citation sends you.
There is also a mode distinction that changes how you should test your own visibility. An answer with no sources was probably not a search answer at all; it may have come from training. If you want to know whether ChatGPT search can find you, you need to trigger an actual search, look for citations and the Sources panel, and repeat the question on different days. A single sighting proves little, because answers vary between runs. What you are looking for is a pattern: does your page appear across repeated, realistic questions, or was one appearance a fluke?
4. Open the door for ChatGPT: OAI-SearchBot and the robots.txt split
Before any writing advice matters, there is a gate, and it is controlled by a few lines in your robots.txt file. OpenAI runs several different crawlers, and confusing them is the single most common technical mistake in this whole topic. The one that matters for citations is OAI-SearchBot. OpenAI’s crawler documentation says it “is used to surface websites in search results in ChatGPT’s search features” and, in plain terms, that “sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links.” If your site blocks it, intentionally or because a security plugin blocked it by default, you have closed the door before the contest starts.
The crucial point is that OAI-SearchBot is not the same as GPTBot. GPTBot crawls content that may be used to train OpenAI’s foundation models. OAI-SearchBot crawls for search. OpenAI states explicitly that each setting is independent: you can allow OAI-SearchBot in order to appear in search results while disallowing GPTBot to indicate that crawled content should not be used for training. Many site owners who object to training have blocked everything with one broad rule and accidentally removed themselves from search answers too. The fix is specific: allow OAI-SearchBot, and make a separate decision about GPTBot on its own merits.
There are two more agents in OpenAI’s documentation, and both are easy to misread in server logs. ChatGPT-User visits a page because a person asked ChatGPT or a custom GPT a question about it; it is not automatic crawling, robots.txt rules may not apply to it, and OpenAI says it is not used to decide whether content may appear in search. Seeing ChatGPT-User in your logs is not evidence that you are in ChatGPT search. OAI-AdsBot only visits pages submitted as ads on ChatGPT and is irrelevant unless you run such ads. When you audit your logs, filter for OAI-SearchBot specifically, and do not mistake a user-triggered fetch for search inclusion.
The practical checklist for this section is short. First, open your robots.txt and confirm there is no rule blocking OAI-SearchBot, including no blanket block that catches it by accident. Second, confirm your host or content delivery network allows traffic from OpenAI’s published searchbot IP ranges, which OpenAI lists at openai.com/searchbot.json, because a firewall rule can undo a correct robots.txt file. Third, remember OpenAI’s timing note: it can take about 24 hours from a robots.txt update for its search systems to adjust, so do not expect instant changes, and do not keep changing the file while waiting. Fourth, verify the crawler is genuine by checking its IP against the published ranges rather than trusting the user-agent string alone, since that string is easy to fake. None of this guarantees a citation. It only makes one possible, which is the necessary first step.
5. Perplexity: numbered citations and PerplexityBot
Perplexity works differently from ChatGPT in one respect that writers should appreciate: its citations are the interface, not an afterthought. Perplexity answers carry numbered citations that link to the original sources behind each claim, so a reader can check any sentence against its source in one click. That design makes Perplexity the most citation-visible of the three engines in this guide, and it also makes it the most unforgiving of vague writing. A claim that cannot be traced to a source stands out, and a source that does not actually support the claim gets found quickly.
The crawler story mirrors OpenAI’s, with the same split between search and training, and the same trap for the careless. Perplexity’s crawler documentation names two agents. PerplexityBot “is designed to surface and link websites in search results on Perplexity” and, notably, “is not used to crawl content for AI foundation models.” Perplexity recommends allowing it in robots.txt and permitting requests from its published IP ranges if you want your site to appear in its search results. The second agent, Perplexity-User, supports user actions: when someone asks Perplexity a question, it may visit a page to help provide an accurate answer and include a link. Perplexity says this fetcher generally ignores robots.txt rules because a user requested the fetch, and it is not used for web crawling or training.
That distinction has two practical consequences. First, blocking PerplexityBot closes the main route into Perplexity’s answers, while blocking Perplexity-User does almost nothing, because user-triggered fetches ignore the rule anyway. Second, as with ChatGPT, a Perplexity-User visit in your logs is not proof of search inclusion; it is proof that one user asked about your page. The signal that matters is whether your pages appear as numbered citations in answers to questions you did not ask yourself.
Perplexity does not publish ranking criteria for which sources it cites, so the same honesty rule applies here as with ChatGPT: no one can sell you its formula, because it has not been disclosed. What is documented is the mechanism: real-time web search, numbered citations to original sources, and a search crawler you must allow. Everything beyond that is inference from observation, and this guide marks it as such wherever it appears. The writing advice in the later sections applies to Perplexity with particular force, because its numbered citations reward exactly the kind of writing that is easiest to verify: a specific claim, in one self-contained passage, backed by a source the reader can open.
6. Google AI Overviews and AI Mode: same index, no separate trick
Google’s position is the simplest of the three, and it is also the one most often misrepresented by sellers of AI-search packages. Google’s guide to optimizing for generative AI features says its AI features are “rooted in our core Search ranking and quality systems.” AI Overviews and AI Mode do not run on a separate index with separate rules. They retrieve from the same Search index, using the same ranking and quality systems, and then generate an answer from what they retrieved. If your page is not doing well in ordinary Google Search, there is no separate AI door to slip through. And if it is doing well, it is already a candidate for AI features.
The eligibility rule Google states is short enough to quote in full: to be eligible to be shown in generative AI features, a page must be indexed and eligible to be shown in Google Search with a snippet, and the site must be included in Search generative AI features in Search Console. Google adds that there are no additional requirements. That sentence does a lot of work. It means the snippet controls you set on a page double as AI controls: if you suppress snippets entirely with nosnippet or max-snippet set to zero, or mark a page noindex, you are also removing it from AI features. It also means the familiar technical requirements carry over directly: Googlebot must be able to crawl the page, the page must return a successful HTTP 200 response, and the page must have indexable content. Google’s May 2025 Search Central post lists exactly those three conditions for content to be considered for AI experiences.
Google also explains when AI Overviews appear at all, and the answer is: only when its systems decide a summary adds something beyond classic results. Many searches never trigger one. That is worth remembering when you measure, because a question that produces no overview is not evidence that your page failed; the feature simply did not fire. AI Mode is described as most helpful for queries needing further exploration, reasoning, or complex comparisons, which is why longer, more structured pages tend to do better there than short answers to simple questions.
There is one more Google point that belongs in this section because it answers a question readers often ask in reverse: can you turn AI Overviews off? From the reader’s side, yes, there are workarounds, and our guide to turning off AI on Google Search walks through the four methods that actually work. From the publisher’s side, the relevant control is the same snippet and indexing controls described above. You cannot opt a page into AI Overviews specifically, and you cannot force it in with a special file or tag, which brings us to the myths section later. For Google, the honest summary is one sentence: do good SEO, keep the page eligible, and you are already in the running.
7. The eligibility checklist: indexed, snippet-eligible, crawlable, working
Before writing a single new paragraph, run this checklist against the pages you want cited. It is the unglamorous half of GEO, and it is where most failures actually live. A brilliant page that returns an error, blocks the wrong crawler, or hides its content behind scripts the retrieval systems cannot read will never be cited, no matter how good the writing is.
First, is the page indexed? Check in Google Search Console for Google, and in Bing Webmaster Tools for the Bing side of ChatGPT’s search partnerships. Indexing is never guaranteed even for a perfect page, but an unindexed page is invisible by definition, so this is the first thing to confirm and the first thing to fix, usually with a sitemap submission and internal links from pages the engines already crawl.
Second, is the page snippet-eligible? For Google, that means indexed and eligible to be shown with a snippet, which rules out nosnippet, max-snippet set to zero, and data-nosnippet applied to the passages you want quoted. For ChatGPT, it means OAI-SearchBot is allowed and its IP ranges pass your firewall. For Perplexity, it means PerplexityBot is allowed and its IP ranges pass. Check all three, not just one, because the engines do not share a gate. A page open to Google but closed to OAI-SearchBot will appear in AI Overviews and never in ChatGPT search answers.
Third, does the page work? Google’s condition is literal: the page must return an HTTP 200 status code. Chains of redirects, soft 404 pages that return 200 with an error message as the content, and pages that collapse to a login wall all fail this test in different ways. A soft 404 is the sneakiest failure, because every automated check says the page loaded fine while the content the engine needed was never there.
Fourth, is the content actually in the page the crawler receives? If your key passages only appear after JavaScript runs, behind a “read more” expansion, or inside an embedded document the crawler does not parse, the retrieval stage may never see them. Google can process JavaScript content when it is not blocked, but every layer of rendering is a chance to fail, and the other engines’ crawlers are simpler. The safest rule is old-fashioned: the sentence you want quoted should be present in the first HTML response, in plain text, where a simple fetch can find it. Test this yourself by fetching the page and searching for the sentence, rather than assuming the rendered page you see in your browser is what the crawler sees.
Fifth, is the page’s canonical signal clean? Duplicate and near-duplicate pages split the engines’ attention and can send a citation to the wrong URL. Pick one canonical version, link to it consistently, and redirect the rest. This is ordinary SEO hygiene, but in an AI context it has a sharper consequence: when the engine does cite you, you want the link to land on the page you actually maintain, not on a stale copy.
8. Write answers machines can quote: direct, self-contained, specific
Once the gates are open, the writing itself decides whether you get cited. The core principle is simple: write the passage the engine would want to lift. An AI answer is assembled from passages, so the passage is the product. A page full of good ideas expressed in long, context-dependent prose is hard to lift safely. A page that states its answer plainly, in one place, in language that survives removal from its surroundings, is easy to lift, and engines prefer what is easy to lift correctly.
Start each section with the answer, not the background. The first sentence of a section should be able to stand alone as the answer to the question the section addresses. Compare two openings for a section about a crawler. “There are several things to understand about how crawlers work, and the history here is interesting” tells the engine nothing it can quote. “OAI-SearchBot is the OpenAI crawler that surfaces websites in ChatGPT search answers; blocking it removes a site from those answers” is a complete, quotable fact. The rest of the section can then explain, qualify, and add detail, but the quotable sentence is already there, waiting at the top.
Make each passage self-contained. Avoid pronouns that point backward (“this means it works”) and references that only make sense if you read the previous paragraph (“as noted above, the second option”). When a passage is lifted into an answer, its surroundings are gone. Names, dates, and quantities should appear inside the passage itself, not two paragraphs earlier. This is not dumbing down; it is writing for a reader who arrives mid-page, which is exactly how an AI answer’s reader arrives.
Be specific where specificity is the value. Vague claims are the easiest to write and the hardest to cite, because they add nothing the model could not already say. “Many businesses benefit from AI search” is not citable; it is a platitude. A specific, sourced statement with a number, a date, and a named source is citable, because it gives the answer something it did not already have. The same rule applies to advice: a step that names the exact setting, file, or menu is quotable; a step that says “optimize your settings” is not.
Finally, answer the question the user actually asked, in the form they asked it. People ask AI assistants in full sentences, with context and follow-ups, not in keyword strings. A page organized around real questions, with headings phrased as questions and answers written in plain language, matches the retrieval stage better than a page organized around keywords. Google says explicitly that its systems understand synonyms and general meaning, so you do not need to repeat every variation of a phrase. Write for the person, and the passage will be ready for the machine.
9. Evidence that earns trust: original data, quotations, sources, dates
Engines do not only ask whether a passage answers the question. They ask whether it can be trusted, and trust, in this context, is mostly a property of evidence. A passage that shows its work is more citable than a passage that merely asserts. This is where the original GEO research is genuinely useful, and where it is often misquoted. The paper that coined the term, “GEO: Generative Engine Optimization” by Aggarwal and colleagues, published at KDD 2024 after appearing on arXiv in November 2023, built a benchmark of diverse queries and tested content-level changes against it. Its abstract reports that GEO methods can boost visibility by up to 40% in generative engine responses, and that effectiveness varies across domains. That “up to” and that “varies” are doing important work: the paper measured a prototype engine the authors built, not the live ChatGPT, Perplexity, or Google systems, and its headline figure is a ceiling in a benchmark, not a promise for your site. Treat it as evidence that content changes can move visibility, not as a forecast of your results.
Within that honest framing, the direction of the research is consistent and useful. The methods that performed best were the ones that added verifiable substance: citing sources, including direct quotations, and adding concrete statistics. The methods that performed worst were the ones that added nothing verifiable: fluency tweaks and keyword stuffing. That pattern matches what the platforms say about themselves. Google’s guide says creating unique, compelling, useful content will likely influence presence in generative AI search more than any other suggestion, and it contrasts first-hand reviews based on personal experience with summaries that simply restate what is already available elsewhere. OpenAI says its ranking aims at relevant, reliable information. All three signals point the same way: originality plus evidence beats polish plus repetition.
In practice, evidence means four habits. First, publish something only you could publish: your own measurements, your own survey, your own test results, your own experience with a product. A first-hand account cannot be replaced by a paraphrase, which is exactly why it gets cited. Second, quote your sources directly and name them, with links, rather than paraphrasing everything into unattributed prose. A quotation with a named source is a gift to a citation system: it is already in the form the answer wants. Third, attach numbers, dates, and names to claims wherever they exist, and keep them current. An undated page making a dated claim is a liability; a page that says when it was checked and what changed since is a reliable source. Fourth, show your author. A named author with visible expertise is a trust signal for readers and for engines; an anonymous page making strong claims is not.
One caution belongs here. Evidence must be real. Invented statistics, fabricated quotations, and fake case studies are not only dishonest; they are increasingly checkable, because the same retrieval systems that cite you also surface the sources that contradict you. A single exposed fabrication can cost a page its credibility across all three engines at once. If you do not have the data, say what you do have, and say what you could not verify. Honesty about limits is itself a trust signal, and it is the one this guide has tried to model throughout.
10. Structure and clarity: headings, lists, tables, clean HTML
Structure is the part of GEO that feels most like ordinary good writing, because it is. Clear headings, short paragraphs, lists, and tables help human readers, and they help retrieval systems for the same reason: they make the boundaries of an idea visible. A page organized into sections with descriptive headings lets a system find the passage that answers a sub-question without reading the whole page, and it lets a human reader do the same. Google’s guide puts it plainly: write content for your human audience, make sure it is well written and easy to follow, and organize it with paragraphs, sections, and headings that provide a clear structure.
Lists and tables deserve special attention because they are the formats AI answers love to borrow. A list of steps, a comparison table, a set of options with their trade-offs: these are already in answer shape. When an engine needs to summarize “the three crawlers and what each does” or “the four methods and how long each lasts,” a clean table on your page is almost ready to be lifted. That is not gaming the system; it is presenting information in the form the question demands. The same logic applies to definitions. A one-sentence definition at the start of a section, followed by explanation, serves both the reader who wants a quick answer and the engine that wants a quotable passage.
Clean HTML matters at a lower level. Use real headings in order, real lists for lists, and real tables for tabular data, rather than formatting everything as styled paragraphs. Semantic HTML is not required by Google, which says it can understand pages that are not perfectly valid, but it helps every kind of reader, including screen readers and retrieval systems, parse the page’s structure. Keep the main content distinguishable from navigation, advertisements, and related-links blocks, because a system that cannot tell your article from your sidebar may quote the wrong one. And keep the page fast and readable on mobile, since page experience is part of Google’s guidance and since many AI answers are consumed on phones.
There is a limit to this advice, and Google states it: there is no ideal page length, and there is no requirement to break content into tiny pieces. Some topics need a long page; some need a short one. Make the page the length its audience needs, and let the structure do the work of making it quotable. A 5,000-word guide with clear sections is more citable than a 500-word page that answers only one question, because the long guide contains more quotable passages for more fan-out queries. But a long page padded with repetition is worse than a short page that says one true thing well. Length is not the strategy; coverage with clarity is.
11. Authority beyond your own page: mentions, reviews, consistency
Everything so far has been about your own page, but engines do not judge a page in isolation. They build a picture of a brand or author from many surfaces: reviews, comparison sites, community discussions, news coverage, directories, and other pages that mention you. If every surface tells the same story, the engine’s picture is sharp and confident, and it is more likely to name you. If the surfaces contradict each other, with different prices, different feature lists, different descriptions of what you do, the picture blurs, and the engine may pick a competitor whose story is consistent, or say nothing at all.
This is why GEO work looks as much like public relations as like SEO. Earning coverage in the publications and communities your buyers trust, encouraging honest reviews, participating in comparisons, and keeping business details accurate everywhere are not side quests. They are how the engines learn who you are. Google’s guide makes the same point from its own side: generative AI responses can include product information and local business details, and keeping Merchant Center feeds and Business Profiles accurate helps those details appear correctly. For a local business, an incomplete or outdated Business Profile is not a minor listing problem; it is a gap in the engine’s knowledge of you.
Consistency deserves emphasis because it is the cheapest authority work available. Check that your name, description, pricing, and key facts match across your own site, your profiles, and the directories that list you. Correct errors at the source, not just on your own page, because the engine may be quoting the directory rather than you. When you find an engine repeating an outdated price, a discontinued feature, or an old review, trace it to the page it is quoting and fix it there. That correction loop, finding the error, fixing it at its source, and re-checking, is one of the highest-value habits in this whole guide, and it costs nothing but attention.
A warning belongs here too. Google’s guide explicitly dismisses seeking inauthentic mentions: its core ranking systems focus on high-quality content while other systems block spam, and generative AI features depend on both. Buying fake reviews, planting false mentions, or manufacturing coverage does not build authority; it builds a record that can collapse. The authority that survives is the boring kind: real customers, real coverage, real consistency, over time. There is no shortcut that the engines have not already seen.
12. What not to do: the myths Google and OpenAI have already answered
Every new discipline collects myths, and GEO has collected an expensive set. The good news is that the platforms have answered most of them directly, so you do not need to guess. Google’s guide has a section bluntly titled “Mythbusting generative AI search: what you don’t need to do,” and it is the single most useful page in this whole topic for anyone being sold a package.
First, you do not need llms.txt or any other special machine-readable file. Google says plainly: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them.” Other engines may read such files for their own purposes, but for Google they are not a requirement, and a vendor who tells you an llms.txt file is the key to AI Overviews is selling you something Google says it ignores. The same applies to special schema.org markup for generative AI search: Google says structured data is not required for it, though it remains useful for ordinary rich results.
Second, you do not need to chop your content into tiny pieces. Google says there is no requirement to break content into tiny pieces for AI to understand it, that its systems can understand the nuance of multiple topics on a page and show the relevant piece, and that there is no ideal page length. Write the page your audience needs, structured well, and stop there.
Third, you do not need to write in a special way just for AI systems, and you do not need to cover every variation of a query. Google says AI systems understand synonyms and general meanings, so you do not need to worry that you have not captured every long-tail variation. Relatedly, creating a separate page for every possible fan-out query, primarily to manipulate rankings or generative AI responses, violates Google’s scaled content abuse spam policy, and Google calls it an ineffective long-term strategy anyway. One strong page beats twenty thin ones aimed at twenty phrasings of the same question.
Fourth, keyword stuffing and fluency tricks do not work. The GEO research found keyword stuffing performed worse than doing nothing, and Google’s spam systems treat it accordingly. Fifth, do not block the crawlers you need while trying to block the ones you do not want. The robots.txt split in sections 4 and 5 exists precisely because a blanket block is the most common self-inflicted wound in this field. And sixth, do not buy citations. No engine sells them, no vendor can guarantee them, and any offer that promises a fixed number of citations for a fixed price is promising something the platforms explicitly refuse to promise. If a pitch cannot show you the platform documentation behind its claim, that is your answer.
13. Measure it: Search Console, prompt panels, and honest reading of the numbers
Work you cannot measure will drift, so the last operational piece is measurement, and here the news is genuinely good. Google now gives publishers a dedicated view of their AI visibility: the Generative AI performance report in Search Console, which covers generative AI features on Google Search and Discover. It shows impressions, meaning times a link to your site was shown inside a generative AI feature, broken down by page, country, device, and date. Our detailed walkthrough, Search Console’s Generative AI Features Report (Beta) Explained, covers each dimension and its quirks, including the fact that chart totals and table totals can differ because of how impressions are aggregated. The key discipline is reading impressions alongside clicks, not instead of them. A page can be cited in many AI answers while its clicks stay flat, because the user got the answer without visiting. Impressions without clicks overstate success; clicks without impressions understate it. Read both.
For ChatGPT and Perplexity there is no equivalent official report, so measurement is manual, and the method matters more than the tool. Build a fixed panel of the questions your buyers actually ask, in their own words, not in keyword form. Run that panel on a schedule, in each engine, and record three things per question: whether you were mentioned, whether you were cited with a link, and whether what the engine said about you was accurate. Repeat over weeks, because a single reading is noise. The trend across the panel is the signal. Our guide to how brands track their visibility inside ChatGPT and Gemini answers explains the metrics in detail: mention share, citation share, sentiment, and the source analysis that turns a tracking exercise into a content list, because the pages your rivals are cited for are the pages you need to beat or outflank.
Three traps will corrupt your numbers if you let them. The first is treating a sample as a census: answers vary by account, location, and day, so one screenshot proves nothing and a trend proves a lot. The second is chasing mentions instead of outcomes: visibility is a means, and the end is what happens when a recommendation turns into a shortlist, a visit, or a sale. The third is testing with your own phrasing and declaring victory. Run the questions your customers ask, including the awkward, misspelled, and indirect versions, because those are the questions the engines actually answer. And check your own analytics for AI referral traffic, including the referral tags the engines attach to outbound links, so you can see which citations actually send visitors rather than merely existing.
A final measurement habit is worth adopting: re-check after every change. When you fix a robots.txt rule, update a page, or correct an error at its source, note the date and re-run the panel two and four weeks later. OpenAI says robots.txt changes can take about a day to take effect for search, and Perplexity says the same, but ranking and citation patterns move more slowly. Without a before-and-after record, you will attribute every change to the last thing you did, which is how myths survive. With one, you will learn which of your actions actually moved the numbers, and that knowledge is worth more than any tool subscription.
14. A 30-day plan — and the honest limits
Here is a plan you can run in thirty days, in order, without buying anything. In week one, do the technical work: confirm your pages are indexed, check the robots.txt rules for OAI-SearchBot, PerplexityBot, and Googlebot separately, confirm your firewall allows the published IP ranges for the search crawlers, and verify that your key pages return HTTP 200 with their quotable sentences present in the first HTML response. Fix any soft 404s and canonical tangles you find. This week alone often explains a site’s invisibility, and it costs nothing.
In week two, pick ten questions your buyers actually ask and write or rewrite one page per question, following the writing rules in sections 8 to 10: answer first, self-contained passages, specific facts with named sources, clear headings, and a list or table where the topic suits one. Do not write twenty pages. Write ten good ones, each aimed at one real question, and link them to each other where the connection genuinely helps a reader.
In week three, do the authority work: check that your facts are consistent across your site, your Business Profile, and your main directory listings; correct the errors you find at their source; and identify the third-party surfaces the engines already cite in your niche, so you know where honest coverage would help most. In week four, build the measurement habit: set up the Search Console Generative AI report, assemble your prompt panel, run it once to establish a baseline, and calendar a repeat run in a month. Then stop, and let the system respond. Citation patterns move on a timescale of weeks, not hours, and frantic daily edits will teach you nothing.
Now the honest limits, because a guide that only promises is not a guide. No tactic in this article guarantees a citation, and no honest vendor can promise one either; OpenAI says placement is not guaranteed, and Google’s features are rooted in ordinary ranking, which is itself never guaranteed. Answers will vary between runs, users, and days, so your visibility will wobble even when you do everything right. Some questions will never produce a citation for anyone, because the engine answers from training or decides no summary is needed. Some niches are dominated by a handful of sources the engines trust deeply, and displacing them takes months of consistent publishing, not one optimized page. And the systems themselves will change: crawler names, report names, and documented behaviors in this guide are current as of 4 November 2026, and they will not all be current a year from now.
None of that makes the work pointless. It makes it ordinary, in the best sense: the same discipline that has always earned trust, applied to a new surface. Be reachable by the engines that matter, write the clearest answer to a real question, show your evidence, keep your facts consistent everywhere, and measure honestly over time. The engines are not looking for a trick. They are looking for the page they can quote without regret. Become that page, and the citations follow as a consequence, not as a prize you chased directly.
0 Comments