Write Clear Claims, Not Robot Food

Somewhere in the last two years, a new worry landed on the desk of every business owner who already had enough to worry about: people are asking ChatGPT, Gemini and Google's AI Overviews questions that they used to type into a search box, and those systems answer in prose. Sometimes they name a business. Usually they don't name yours.

The reflex response is to ask what the machines want and then feed them. That reflex is being monetised fast. There are agencies in India selling "GEO packages" and "LLM optimisation" that amount to stuffing pages with question-shaped headings, adding seventeen kinds of schema markup, and writing in a flat, hedged register that nobody enjoys reading. Some of it is harmless. Some of it actively makes your pages worse for the humans who were going to buy from you anyway.

The honest position is this: nobody outside the companies building these systems knows precisely how they select what to cite, the selection changes without notice, and anyone claiming certainty is guessing confidently. But we can say something useful anyway, because the things that make a page easy for a language model to quote accurately are, almost entirely, the things that make it easy for a busy person to read and trust.

What a machine can and cannot lift

A large language model answering a question about tax filing deadlines or AMC rates for lift maintenance is doing something like assembling a short answer from text it has read, and — in the versions that browse live — from pages it has just fetched. For it to attribute a claim to you, the claim has to be liftable: a complete thought in a small number of words, that survives being removed from everything around it.

Here is a sentence that cannot be lifted:

Pricing depends on a number of factors, and we work closely with each client to understand their unique requirements before arriving at a figure that works for everyone.

There is nothing in it. A model cannot quote it, a customer cannot act on it, and a competitor's page that says "A single-location restaurant usually needs two POS terminals; most of the cost sits in the terminals, not the software licence" will be quoted instead. Not because that page is better optimised. Because it said something.

So the first edit on any page you want cited is a search-and-destroy mission against sentences that survive deletion. Read each one and ask: if I cut this, does the page lose information? If not, it was decoration.

The shape of a liftable claim

• One idea, one sentence. Not three clauses joined by "and also".
• Self-contained. If the sentence starts with "This means" or "As mentioned above", it cannot travel. Name the thing again.
• Specific enough to be wrong. "Fast loading" is unfalsifiable. "Google's Core Web Vitals treat a Largest Contentful Paint above 2.5 seconds as needing improvement, and above 4 seconds as poor" is a checkable statement that someone could contradict.
• Scoped. "For a GST-registered trading business in Delhi" is better than nothing, because it tells the reader and the machine when the claim applies.

Attribution is the whole game

The thing these systems are most obviously trying to avoid is stating something false and being blamed for it. That shapes what they are comfortable repeating. A number with a visible source and a date attached is safer to repeat than a number floating on its own.

Which gives you a practical rule that costs nothing: every fact on your page should carry its origin and its age in the sentence itself. Not in a footnote, not in a citation block at the bottom. In the sentence.

Compare:

• "The deadline is 31 July." — Deadline for what, which year, says who?
• "For individual taxpayers whose accounts do not require audit, the income tax return due date has been 31 July of the assessment year; check the Income Tax Department's current notification, as the date has been extended in several recent years." — Scoped, attributed, honest about volatility.

The second version is longer and less confident, and it is the one that gets quoted, because it is the one a cautious system can repeat without risk. It also happens to be the version that does not get your reader into trouble.

This is where most "AI-optimised" content fails. It is confident and sourceless. It reads like it was generated, because it was. If your page about commercial kitchen equipment says "most restaurants replace their chimneys every five years" with no basis, you have added a liability, not an asset. Say "in our experience fitting kitchens in East Delhi, chimney filters need replacing far more often than the chimney itself" — clearly marked as your own observation — and it becomes attributable to you, which is the point.

Structure that isn't theatre

Structured data — the invisible code that tells search engines "this is a product, this is its price, these are its opening hours" — is worth having for the things it genuinely describes. Organisation details, local business address and hours, product price and availability, FAQ entries that really are questions. It is cheap, it is documented by Google, and it removes ambiguity.

What it is not is a ranking lever you can pull harder. Marking up a page with Article, FAQPage, HowTo and Speakable schema when the page is a 300-word brochure does nothing except make an audit tool show green. If an agency's AI visibility proposal is mostly a schema checklist, ask what the schema is describing. If the answer is thin, the schema is thin.

The structure that actually matters is the visible kind:

1. Headings that state the answer, not the topic. "How long does a GST registration take?" is a topic. "GST registration usually takes a few working days once documents are clean" is a heading that answers before you read the paragraph.
2. The answer first, the reasoning after. Journalists call it the inverted pyramid. Put the conclusion in the first two sentences under each heading. Everything a model is likely to extract sits in that opening.
3. Tables for anything comparative. Price tiers, turnaround times, what's included at each level. A table is unambiguous in a way that a paragraph describing three options is not.
4. A visible last-updated date on anything that can go stale.

None of that is for robots. It is what a reader who has four tabs open and six minutes does with a page anyway: scans headings, reads the first line of the section that matches their problem, leaves.

What not to buy

Some things being sold in this category right now that we would not spend money on:

• Question-stuffed FAQ blocks. Thirty questions appended to every page, each answered in a sentence that repeats the page title. It's padding. Three real questions beat thirty fake ones.
• "Guaranteed AI Overview placement." Nobody controls this. Google does not sell it and neither can an agency.
• Entity spamming. Mentioning your brand name in the third person forty times so the model "learns the entity". It reads as badly as it sounds.
• Rewriting your whole site into Q&A format. Service pages have a job: explain what you do and make it easy to contact you. A service page written as an interview with nobody does that job worse.
• AI-written bulk content at volume. Pages that confidently assert things with no source are precisely the pages these systems have reason to discount.

And one thing worth checking before any of this: whether the crawlers can reach you at all. Content strategy is irrelevant if your robots file blocks GPTBot and your server returns errors to anything without a browser user agent. That's a separate, mechanical question, and it's worth settling first.

An honest note on what we don't know

We do not know how heavily any given system weights freshness against authority. We do not know whether being cited in an AI answer sends meaningful traffic for most businesses or almost none — the measurement is poor and the referrer data is patchy. We do not know how long today's behaviour lasts.

What we do know is that the cost of writing clearly, sourcing your claims and structuring your pages honestly is near zero if you were writing anything at all, and the payoff exists regardless of which way the AI question lands. A page that states a real claim, attributes a real fact and answers before it explains is a better page in 2019 terms, 2026 terms, and in the terms of the customer reading it on a phone in a traffic jam on the Noida link road.

That is the whole argument. There is no robot dialect to learn.

What to do this week

Pick your three most commercially important pages. On each one, do four things: delete every sentence that survives deletion; add a source and a date to every factual claim; rewrite each heading so it states the answer; and move the conclusion of each section to its first line. That is an afternoon's work and it needs no tools.

If you want a second pair of eyes on whether your pages are saying anything liftable — or you want the crawler-access side checked properly — our AI visibility work and our SEO practice start from the same place: read the page as a customer would, then fix what's actually broken. Send us a page you're unhappy with and we'll tell you what we'd cut.

Questions we get asked