Answer engine optimisation: getting quoted by AI assistants

What changed about how buyers search

A buyer looking for a software vendor used to type a phrase into Google, open four results and compare them. A growing share of that same research now happens in a conversation with an assistant, where the person asks a question in full sentences and receives a synthesised answer with a handful of citations underneath.

The commercial consequence is straightforward. In the old model you competed for a position on a page of ten links. In the new one you compete to be one of perhaps three sources the assistant draws from. Being ranked eighth for a term still brought some traffic. Being the eighth most quotable source brings none.

None of this makes traditional search irrelevant. Both systems read the same web pages, and the work that makes a page quotable also tends to make it rank. The framing is different, though: you are writing for a reader who will extract a paragraph rather than browse your site.

How an answer engine decides what to quote

The exact selection methods are not published and they change. The observable behaviour is consistent enough to work with.

The page has to be retrievable. Assistants either search an index or crawl live. A page blocked from crawling, rendered only by client side JavaScript, or buried where nothing links to it, is not a candidate regardless of quality.

The answer has to be extractable. A model assembling a response needs a passage that stands on its own. A paragraph that begins "as mentioned above" or depends on the previous three sections cannot be lifted cleanly, so a competitor's cleaner paragraph is used instead.

The claim has to be specific. Between "we have extensive experience in manufacturing software" and "we have delivered more than 200 projects since 2017, including work for PPAP and Kokuyo India", the second is quotable because it contains checkable facts. The first says nothing a model can repeat usefully.

The source has to look accountable. Content attributed to a named person with a stated role and credential is treated more seriously than anonymous marketing copy. This is the same principle behind Google's guidance on experience, expertise, authoritativeness and trust, applied by a different consumer.

Seven changes that make a page quotable

1. Answer in the first paragraph

Open every important page with a direct answer of roughly forty to sixty words that could be lifted verbatim. Name the business, what it does and where it operates inside that block. Then continue with context. The common failure is a warm introductory paragraph that delays the answer until halfway down.

2. Use headings shaped like questions

Match how people ask. "How much does custom software development cost in India" works better than "Pricing". Then answer immediately in the first sentence under the heading rather than building to it.

3. Keep paragraphs to one idea

One to three sentences each. Models extract paragraphs, and a paragraph carrying four ideas is harder to quote for any of them.

4. Replace adjectives with figures

Specific numbers, timeframes and mechanisms survive extraction. "Results in six to eight weeks" can be quoted. "Fast results" cannot.

5. Add a real FAQ section

Three to five questions buyers actually ask, each answered completely in two to four sentences without depending on the rest of the page. These are the passages most often lifted.

6. Mark it up with structured data

Organization, Service, FAQPage and Article markup tells a machine what the page is about. The text inside the markup must match what a visitor sees. Google's structured data guidelines are explicit that marked up content should be visible on the page, and mismatches risk losing rich results entirely.

7. Name the author

A byline with a real person, their role and their relevant experience. Anonymous content is harder to trust and harder to attribute.

A useful test: read one paragraph of your page aloud with no surrounding context. If it makes complete sense and contains a checkable fact, it is quotable. If it needs the paragraph before it, rewrite it.

The technical floor you still have to clear

Answer engine work sits on top of technical health rather than replacing it. If a page cannot be crawled or takes too long to become usable, nothing else matters.

Google publishes Core Web Vitals thresholds that define a good experience: Largest Contentful Paint within 2.5 seconds, Interaction to Next Paint within 200 milliseconds, and Cumulative Layout Shift below 0.1. These are measurable on your own pages rather than matters of opinion, which makes them a sensible starting point for any audit.

Beyond speed, the recurring technical problems are unremarkable and common: important content rendered only by JavaScript, several pages competing for the same term so none of them wins, missing or duplicated canonical tags, and thin pages generated in bulk that dilute the site. Fixing these is unglamorous and usually produces more improvement than any content change.

Why attribution decides whether you are trusted

Both Google's systems and AI assistants have the same underlying problem: distinguishing content written by someone who has done the work from content assembled by someone who has not. Neither can verify expertise directly, so both rely on signals.

What helps is straightforward. Attribute pages to a named person with a role and a stated basis for their expertise. State facts that could be checked, such as when the company was founded, how many projects it has delivered, and which clients it has worked with. Keep the business description consistent across your site, your other domains and your directory listings, so an entity can be resolved reliably.

What hurts is equally straightforward, and it is worth being blunt because it is widely sold in this market. Invented review markup, fabricated testimonials and claimed credentials nobody holds are violations of Google's spam and structured data policies. Sites carrying them can lose rich results or rankings. The risk lands on your domain and stays there after the agency that suggested it has moved on.

How to measure something you cannot rank in

There is no equivalent of a rank tracker for AI answers, and anyone selling a precise ranking position inside ChatGPT is selling an estimate. Four things can be measured honestly.

Ask the assistants directly. Keep a list of the twenty questions a buyer would ask before hiring you. Run them through the major assistants on a schedule and record whether you are cited. Crude, repeatable, and more informative than any dashboard.

Watch referral traffic from assistant domains. Citations that get clicked appear in analytics as referrals. This undercounts, because many people read the answer without clicking, but the trend is real.

Watch server logs for AI crawlers. They identify themselves. If they are not fetching your pages, that is the problem to fix first.

Keep tracking ordinary search. Impressions and clicks in Search Console still describe most commercial discovery, and the work overlaps heavily.

Mistakes that get pages ignored or penalised

Publishing volume instead of answers. Pages of general commentary lose to a smaller number of pages that answer real buying questions completely.

Burying the answer. Long introductions before the point are a common problem on service pages.

Structured data that contradicts the page. FAQ markup with questions that do not appear in the visible text is a policy violation and can cost you rich results.

Facts nobody can attribute. Statistics with no source, and claims like "trusted by leading brands" with no names, give a model nothing to quote.

Treating this as a separate budget. Nearly everything described here also improves conventional rankings. An agency selling answer engine optimisation as a distinct product on top of your existing SEO retainer is usually selling the same work twice.

Where to start if you only have a week

Most companies reading this will not rebuild their site. Five changes carry most of the benefit and can be done on existing pages without a redesign.

Day one, fix the opening of your four most commercial pages. Put a forty to sixty word direct answer at the top of each one, naming the company, what it does and where it operates. This edit usually returns the most for the effort, because it is the passage most likely to be extracted and it is currently a welcome message on most sites.

Day two, write the questions down. List the twenty things buyers ask you before they sign. Sales calls and your inbox are the source, not keyword tools. Those questions are your content plan and your FAQ sections, and they are phrased the way real people phrase them.

Day three, add FAQ sections to the commercial pages. Three to five of those questions per page, answered completely in two to four sentences each, with matching FAQPage markup that repeats the visible text exactly.

Day four, add authorship. A byline with a real name, role and relevant experience on every page that makes a claim, plus Person and Organization markup connecting them.

Day five, measure the technical floor. Run the commercial pages through a Core Web Vitals check and confirm they are crawlable and indexed. If they are not being fetched, everything above is theoretical.

None of this requires new software or a larger budget. It requires deciding what you actually claim and then saying it in the first paragraph.

Questions buyers ask

Sources and basis for this article

• Google's published Core Web Vitals thresholds for Largest Contentful Paint, Interaction to Next Paint and Cumulative Layout Shift, documented at web.dev and in Google Search Central documentation.
• Google Search Central structured data guidelines, which require marked up content to be visible on the page, and Google's spam policies covering fabricated review markup.
• First hand experience at Okay Web Pvt Ltd since 2017, across more than 200 delivered projects including SEO and content work.

If you want to know whether your pages are structured to be quoted, send the domain to ratna@oakyweb.com, or read more about how the SEO work is run.