Being cited is not the same as ranking, and the difference decides how you write. A ranked page wins a click. A cited page has one of its sentences lifted into an answer, often with no click at all, which means the unit of work is the passage rather than the post. There is now peer-reviewed evidence about which passages get lifted, and some of it contradicts twenty years of habit, so generative AI optimisation has become a question of evidence rather than opinion.
Key Highlights
- Quotations and statistics carry the weight: a study done at KDD 2024 showed that quotes increase visibility by about 41%, while statistics increase visibility by about 31%.
- Keyword stuffing decreased visibility, according to the same study, by about 8%, which is the opposite of what was believed before.
- Google requires no special markup to be cited: a page needs to be indexed and snippet-eligible, and the documentation says there are no additional technical requirements.
- Blocking OAI-SearchBot removes you from ChatGPT search answers, and OpenAI states robots.txt changes take about 24 hours to register.
- Manual fact-checking is now written guidance, extended since 1 October 2026 to cover titles, meta descriptions, structured data and image alt text.
What Makes a Passage Quotable?
Three properties, and none of them is keyword density. A quotable passage is self-contained, so it survives being removed from the page. It is attributable, meaning the claim inside it can be traced to a source a model can name. And it is specific, because a sentence carrying a figure, a date or a named constraint is more useful in an answer than a sentence carrying an adjective.
This is not the same as writing for rankings. A ranking article can make its point through eight hundred words and reward the person who takes the time to read it. An article meant to be cited has to answer the question in a couple of sentences and build the argument beneath it since the extraction happens before the reader gets there. While most advice about AEO marketing goes only as far as answering the question (which is relatively easy), the harder part is in making the answers self-contained.
There is a term used in the academic community for what is being measured. The study that led most of the research done on this measures the visibility of citations by the position-adjusted number of words, which means counting the words in sentences containing a citation based on their placement within the answer. What this suggests is worth contemplating: it’s not your presence but your share of the answer.
How Do the Nine Tactics Rank on Evidence?
The following ranking is based on the main metric of the team from IIT Delhi and Princeton who evaluated nine content changes over around 10,000 queries and presented their findings at KDD 2024. Baseline is 19.3.
| Tactic | Effect on visibility |
| Adding quotations from credible sources | Around 41% increase |
| Adding relevant statistics | Around 31% increase |
| Improving fluency of the writing | Around 28% increase |
| Citing sources explicitly | Around 27% increase |
| Adding appropriate technical terms | Around 18% increase |
| Making the text easier to understand | Around 14% increase |
| Writing in a more authoritative voice | Around 10% increase |
| Adding unique words | No meaningful gain |
| Keyword stuffing | Around 8% decrease |
Four of these group into a single strategy: say something that can be verified and provide a source. Quotations, statistics, and citations all fall under the same strategy, and these contribute to the three largest increases in the table. The third place ranking for fluency is a difficult result for any author producing work in volume, as fluency is the first casualty of an unedited draft, and any GEO AI citation tracking will know that the four strategies above will increase results before the others.
The two strategies that have failed to produce results provide their own insight. The lack of unique words did nothing for an author’s success, which is evidence against the theory that the reward comes in the form of distinctive vocabulary usage. Any GEO AI tool that scores your draft on keyword density is optimising against the only published evidence available.
Why Keyword Stuffing Now Costs Visibility
The mechanism is not mysterious. A generative engine is selecting passages to support a claim it is already assembling, so it reaches for text that reads as a source rather than text that reads as a target. A paragraph carrying its keyword six times reads as the latter, and the paper’s authors note explicitly that classic search-engine-optimisation tactics of this kind perform poorly in their tests.
The Google argument is similar but from a slightly different angle. The spam policies of Google address content scaling as pages that are created for manipulation purposes rather than user assistance, regardless of whether automation, human efforts, or a combination of both are used; there is no publicly posted threshold for such content. As a result, a stuffed page faces double jeopardy: its performance suffers while it falls closer to the line than an unstuffed page.
So the first thing to cut from an AEO marketing brief is the density target. That is not an argument for removing the phrase people actually search. It is an argument for using it where it belongs, once or twice per section, and spending the remaining effort on the four tactics that measurably work.
The Princeton Caveat Worth Stating
Any honest use of that study needs its limits attached, because they are substantial and the figures circulate without them. The experiments ran on a simplified two-step generative engine built on GPT-3.5-turbo over the top five Google results, not on the live commercial systems people actually use. The subjective half of the scoring was performed by GPT-3.5 under a G-Eval-style method, and the authors report those scores were poorly calibrated and had to be normalised. Only single-turn engines were evaluated.
In addition, the efficiency depends on the domain, and the researchers advocate for domain optimisation, meaning that the 41% number is the average of twenty-five different domains, not a promise for yours. Consider the ranking to be an order, but not necessarily correct magnitudes, and set expectations in your AI citation tracking reports accordingly.
That is a more useful way to hold it than either dismissing it or quoting the headline number as settled. The ordering has held up against everything Google has since published about helpful content, and the ordering is what you act on.
Writing the Answer First
The practical shape follows from all of the above. Open each section with the answer in one or two sentences, stated plainly enough to be quoted without the surrounding paragraph. Then give the reasoning, the caveats and the examples beneath it, which is where a reader who clicked through gets their value.
Keep one claim per sentence. A sentence carrying three claims cannot be extracted cleanly, so it tends not to be, and the habit of stacking qualifiers is the most common reason a well-researched paragraph never gets quoted. Attach a figure, a date or a named source to the claims that matter, since that is the quotation and statistics finding applied rather than admired.
Then edit for fluency, which the evidence puts third and most workflows put last. Reading a draft aloud catches the clauses a model would stumble over, and it is the cheapest quality step in generative AI optimisation by a wide margin.
Where Google Finds Its Citations
The eligibility rule is short. A page must be indexed and eligible to appear with a snippet, meeting Google’s technical requirements, and the documentation states plainly that there are no additional technical requirements. No machine-readable file, no AI-specific markup and no schema type opens a door here, and Google says so explicitly.
What does change the odds is how many chances a page gets. AI Overviews and AI Mode issue several related searches across subtopics while assembling a response, which Google calls query fan-out, and the result is a wider and more diverse set of links than classic search returns. A site with genuine depth across a subject has more surfaces exposed to those fan-out queries than a site with one page on the topic, which is the strongest argument for topical clusters that has emerged in years. No GEO AI platform can manufacture that depth for you; it can only report its absence.
Two further details belong in any plan. AI Overviews appear only when Google judges them additive, so they often do not trigger at all, and AI Mode may use different models from AI Overviews, so their links differ. Google has also said it sees clicks from pages carrying AI Overviews as higher quality, with users more likely to spend time on the site, which is the counterweight to the assumption that citation always cannibalises the visit.
Crawler Access, Bot by Bot
Access is where most citation problems actually live, and the settings are independent of each other. OpenAI documents four crawlers and the consequences of blocking them.
| Crawler | Purpose | Effect of blocking |
| OAI-SearchBot | Surfaces sites in ChatGPT search features | Opted-out sites will not be shown in ChatGPT search answers, though they can still appear as navigational links |
| GPTBot | Crawls content that may train OpenAI’s foundation models | Signals that content should not be used for training; separate from search |
| ChatGPT-User | Fetches pages when a user asks ChatGPT or a Custom GPT a question | Not used for automatic crawling, and OpenAI states it does not determine search eligibility |
| OAI-AdsBot | Checks safety of landing pages submitted as ChatGPT ads | Visits only submitted ad pages |
Four operational points come with that table. To appear in ChatGPT search, OpenAI recommends allowing OAI-SearchBot in robots.txt and allowing its published IP ranges. Changes to robots.txt take about 24 hours to register for search. Allowing both OAI-SearchBot and GPTBot means one crawl may serve both purposes. And the full user-agent strings carry version numbers that change, so a rule written against an exact version will rot.
On Google’s side the distinction to hold is between robots.txt for Googlebot, which governs crawling for Search including its AI features, and Google-Extended, which limits training and grounding in some of Google’s other systems rather than acting as a control over AI features in Search. Conflating the two has cost sites their eligibility, and the mistake is usually made at the content delivery network rather than in the file.
Can a Model Find Your Best Paragraph?
Often not, and the reasons are mundane. Snippet directives are the first: nosnippet, data-nosnippet, max-snippet and noindex all limit what Search can show from a page, including inside AI features, and a restrictive max-snippet inherited from an old build quietly caps your share of every answer.
Text is the second. Google’s guidance asks that important content be kept in text, supported by images and video where relevant, which rules out the practice of putting the specification table or the key figure inside a graphic. A number in an image cannot be quoted, and a number in a caption usually is not.
Structured data is the third, and the rule is narrower than most templates assume: it should match the visible page text. Markup that asserts something the page does not say is a violation rather than an advantage, and it is a trust problem on top. None of this makes a page citable on its own, which is the honest limit of technical work. It removes the obstacles, and the writing does the rest.
Which Five Faults Block a Citation?
- Buried answers: The usable sentence sits in paragraph six, under three paragraphs of context that an extraction step never reaches.
- Unsourced claims: A figure with no attribution is the one thing a careful engine will not quote, and attribution is the cheapest fix on this list.
- Stacked clauses: Three claims in one sentence cannot be lifted cleanly, so the paragraph is passed over in favour of a thinner but tidier one elsewhere.
- Key facts in images: Specifications, prices and comparison tables rendered as graphics are invisible to the thing deciding what to quote.
- Inherited restrictions: A nosnippet or tight max-snippet left over from a scraping scare, or an AI crawler blocked at the network edge rather than in robots.txt, where nobody thinks to look.
Four of the five are editing faults rather than technical ones, which is why generative AI optimisation belongs with the editors rather than filed under development, and why AEO marketing that arrives as a technical audit alone leaves most of the gain untouched.
AI Citation Tracking on a Budget
Start with a fixed prompt panel, because it is the only measurement that answers the question you actually have. Twenty to thirty buying and research questions, written the way customers phrase them, run monthly across the assistants your market uses, scored on whether you appear, whether the claim attributed to you is accurate, and which competitor is named instead. Thirty prompts read carefully beat a dashboard reporting one number.
Server logs give the hard evidence that no third-party tool can. Filter for OAI-SearchBot, GPTBot, ChatGPT-User and Googlebot, and compare what arrived against what was served, because that is how you learn an access change took effect rather than hoping it did. Referral traffic from assistant domains in your analytics completes the picture, small in volume and usually late in the buying process.
Search Console has one limitation that shapes every report built on it: appearances in AI Overviews and AI Mode are folded into overall search traffic under the Web search type rather than broken out, so treat movement as a direction. Anyone selling AI citation tracking as a precise measurement of Google’s AI surfaces is selling an estimate, and a tool that scores several assistants into a single visibility index is compressing different systems with different models into one number. Useful as a trend, misleading as a diagnosis.
What Google Asks You to Check
One instruction, added to the accuracy section of Google’s generative AI content guidance on 1 October 2026, changes the workflow rather than the targeting: it is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing. The guidance explains that generative models predict likely sequences of words rather than retrieve facts, which is why their output can contain errors.
The scope is the part teams miss. That review covers title elements, meta descriptions, structured data and image alt text, which are precisely the artefacts generated in bulk and checked least. Google’s position since February 2023, that appropriate use of AI is not itself a violation, is unchanged; what has changed is that a workflow unable to show a human checked the claims is now operating against written guidance.
Two smaller notes close the loop. A documentation update on 15 May 2026 clarified that Google’s spam policies apply to generative AI responses in Search, including AI Overviews and AI Mode. And Google’s guidance on optimising for generative AI features dismisses the shortcuts directly, naming AI-specific text files, mechanical chunking, AI rewriting for its own sake and inauthentic mentions as unhelpful, while calling the discipline itself still SEO. Worth remembering before any GEO AI vendor proposes a file for you to publish.
Savit’s Test for a Citable Page
At Savit Interactive, we read a draft by covering everything except one paragraph and asking whether that paragraph would survive on its own in someone else’s answer. If it needs the paragraph above it to make sense, it is not citable yet. Our editors apply the evidence-backed order rather than the fashionable one: a verifiable claim, a named source, then a fluency pass, with keyword placement last and deliberately light. Having edited search content from Mumbai since 2004, we have learned that the pages which get quoted are the ones a subject expert would recognise as correct, whatever the sector. Firms comparing proposals should ask who fact-checks, and ask to see the fixed prompt panel rather than the dashboard screenshot, since AI citation tracking is only as good as the questions in it.
