How Pages Get Picked Up by ChatGPT, Perplexity and AI Overviews

What actually happens between the question and the citation

Most advice on this subject skips the mechanism, which is why it ends up as "write helpful content" with extra steps. The mechanism is knowable and it dictates almost every practical decision below.

A question arrives. The system rewrites it into several sub-questions, because a real question like "which SEO agency should a Bangalore SaaS company hire" contains four separate lookups. Each sub-question hits an index — Google's own for AI Overviews, a mix of licensed and self-crawled indexes for the assistants. Candidate passages come back. The model ranks them, drops the ones that repeat each other, writes an answer from the survivors, and attaches citations to the passages it actually used.

A question arrives. The system rewrites it into several sub-questions, because a real question like "which SEO agency should a Bangalore SaaS company hire" contains four separate lookups. Each sub-question hits an index — Google's own for AI Overviews, a mix of licensed and self-crawled indexes for the assistants. Candidate passages come back, the model ranks them, drops the ones that repeat each other, writes an answer from the survivors, and attaches citations to the passages it actually used.

Write claims that survive being cut out of the page

This is the change with the biggest effect and the lowest cost, and it's an editing discipline rather than a technical one. Assume every paragraph you write will be read alone, by a machine, with no title, no heading and no surrounding context. Then write accordingly.

Six rules cover most of it.

1. Restate the subject instead of using pronouns. "It usually takes three to six months" is unusable on its own. "SEO usually takes three to six months to show movement on a competitive keyword set" is a complete, liftable claim.
2. Put the claim in the first sentence, the reasoning after. Inverted pyramid, but at paragraph level rather than article level. A retriever scoring a chunk sees the opening first.
3. Attach a number, a range or a date. "Fairly expensive" gets skipped. "₹75,000 a month, ex-GST" gets quoted. This is also the finding of the research that named this field: sources with statistics, quotations and citations got picked up more often than sources without.
4. Keep the answering unit to roughly 40–80 words. Longer, and the retriever has to decide which half of your paragraph is the answer — and it may decide wrong, or decide to use someone shorter.
5. One idea per paragraph, one question per heading. Headings that match how people actually ask are free retrieval signals. Headings like "Our Approach" match nothing.
6. Date the claims that expire. "As of July 2026" costs three words and tells a system with a mixed-vintage index which version of a fact yours is.

Be one entity, everywhere, in the same words

Ranking tolerates inconsistency. Being described accurately does not. When an assistant is asked what your company does, it assembles an answer from whatever it can retrieve about you — your site, your LinkedIn, a directory listing from 2021, a press mention that got your founding year wrong. It will pick a version. There is no guarantee it picks yours.

Entity consistency is a finishable project, which makes it unusually satisfying work in a discipline where nothing is ever finished. A focused week gets most of it done.

• One canonical name string. Decide whether you're the legal entity or the trading name and use that form in every description. Machines resolving entities are matching strings before they're matching meaning.
• One one-sentence description, reused verbatim on your homepage, About page, LinkedIn, Crunchbase, directory listings, conference bios, YouTube channel and press boilerplate. Verbatim, not "broadly similar".
• Consistent name, address and phone if you have a physical presence — including formatting. "Ground Floor" and "G/F" are different strings to a matcher.
• Real named authors with real bio pages. A byline that goes nowhere is not an author entity. The bio page should state credentials specifically and link out to that person's profiles elsewhere.
• `Organization` schema with `sameAs` pointing at every profile you just made consistent. This is the machine-readable version of the same statement.
• Go and kill the old descriptions. The 2021 directory listing calling you a "web design studio" is retrievable and will be repeated. Update it or get it removed.

You can't be your own only source

Here's the part that no amount of on-page work substitutes for. Retrieval systems weight agreement across sources. A claim only you make, on a domain only you control, is one source saying one thing. Three independent sources saying it makes it a fact the model is comfortable stating.

This is also where brand mentions behave differently from backlinks. A link is a vote that passes ranking signals. A mention on a page these systems retrieve from can earn you a citation with no link at all, because what's being read is the text, not the anchor. Unlinked coverage that classic SEO reporting ignores genuinely counts here.

Structured data: what helps, and what's sold as helping

Google has been explicit that there's no special markup that gets content into AI Overviews, and no assistant has published a markup requirement. That doesn't make schema pointless — it makes it hygiene rather than a lever, and it's worth knowing which parts do real work.

Crawler access is the gate nobody checks

This is the one that quietly cancels everything else, and it usually gets broken without a search person in the room. Somebody reads about AI training on copyrighted work, adds a blanket disallow, and removes the company from every citation surface while doing nothing whatsoever about content already trained on. The bots are separate and their jobs are different. Decide per bot, in writing.

Tracking citations without paying for a tool

The tools in this category are improving and some are worth the money at scale. You do not need one to start, and building the manual version first means you'll know what the tool is actually measuring when you buy it.

1. Freeze a prompt set. Thirty to fifty questions your actual buyers ask, written down before you change anything. Real phrasing, not keyword phrasing — "is it worth hiring an SEO agency for a 20-person SaaS company" rather than "seo agency india".
2. Run each prompt three times per engine, monthly. Answers vary between runs; the same question can return different sources twice in a row. Log a citation *rate*, not a yes/no. A boolean is the single commonest flaw in AI-visibility reporting.
3. Use a clean, logged-out browser with the right country set. Personalisation and location change the answers, and testing from your own account measures your own history.
4. Log five columns: prompt, engine, cited (out of three runs), which URL, and whether the description of you was accurate. That last column earns its place — misdescription is fixable and invisible if you only count citations.
5. Build a referral segment in GA4 for chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com. ChatGPT also appends utm_source=chatgpt.com to outbound links, which makes it easy to isolate. Note that AI Overview clicks arrive as ordinary Google organic and can't be separated this way.
6. Grep your server logs for the crawler user agents in the table above. Free, immediate, and the earliest possible signal — if OAI-SearchBot has never fetched you, no amount of paragraph editing is the problem.

Five things we'd stop doing tomorrow

The response to this shift has produced as much waste as it has useful work. These five come up constantly.

• Stop bolting FAQ blocks onto the end of every page. Six questions that restate the article in interrogative form add length, not answers. If a question deserves an answer, it deserves a heading and a real paragraph.
• Stop publishing `llms.txt` and calling it a strategy. It costs nothing, so publish it if you like. Don't put it on a scope sheet, and don't let anyone charge you for it.
• Stop writing 400 words of context before the answer. It was already bad for readers. Now it means the first retrievable chunk of your page is a preamble.
• Stop putting numbers in images and PDFs. A price in a JPEG is a price that can't be quoted, compared or cited. Same for spec tables rendered as screenshots.
• Stop treating citation count as the KPI. It's a leading indicator at best. We've seen the logic used to defend a retainer where mentions climbed steadily and enquiries didn't move at all.

What a citation is worth, and what we still measure instead

Everything above is worth doing, and we do it inside a normal SEO engagement rather than as a separate line item — partly because the same edits also help you rank, and partly because we don't think citation counts are an honest thing to be accountable for. The same prompt returns different sources on different runs. Building a guarantee on a number that fluctuates that much would be a nice way to look confident while promising nothing.

So the accountability stays where it's always been. We freeze your trailing-90-day qualified organic lead count on day one and guarantee movement against that number. Miss it in 90 days and we keep working free until we beat it. We never promise a specific ranking position for a specific keyword, and we won't promise a specific citation either. SEO runs from ₹75,000/mo, or ₹40,000/mo for smaller sites, ex-GST, month-to-month after the first quarter, and you keep every asset if you leave. It's why we take three clients a month — a guarantee needs slack in the roster to be worth anything.

If you're doing this yourself, the order that matters is: crawler access first, because it gates everything. Then entity consistency, because it's finishable and it stops a wrong description spreading. Then the paragraph-level editing, which is cheap and compounds. Then earned media, which is slow, expensive, and the part that actually makes you a source.

And before you spend a rupee on any of it, work out how much you've genuinely lost. Segment your Search Console data first. It takes an afternoon and it decides whether this is a project or a paragraph in next quarter's plan.