serps.io
Guides
Intermediate·6 min read

Getting cited by AI answer engines

What generative engine optimization actually involves: choosing prompts, writing passages a model can lift, and tracking who gets cited.

What changes when the answer replaces the list

Classic search returned ten links and let the reader choose. An AI answer composes a response and names a few sources, which collapses the payoff curve: being the source quoted is worth a great deal, and being the eleventh relevant page is worth nothing at all. Two things follow. The unit of optimization stops being the keyword and becomes the question, since people type full sentences into a chat box and fragments into a search box. And the content that wins is whatever is easiest to lift and attribute, not whatever is longest. What does not change is most of the work: these systems retrieve from an index, so a page that is not crawlable, not indexed, or not relevant will not be cited either. Treat this as an extra surface on top of your existing search work, and be suspicious of advice that says to discard the fundamentals. It also means one page can serve both surfaces. A page that answers its question clearly in the opening lines tends to do better in conventional results too, because the people who click it find what they came for and stop searching.

Pick the prompts you want to win

Start by writing down the twenty to fifty prompts a buyer would actually type, as full sentences. Not best crm, but which crm works for a two-person agency that bills hourly. Get them from sales calls and support conversations, since your team already hears them daily. Group them three ways: prompts where you should obviously be named, such as your own category and your own comparisons, prompts where being named is plausible, and prompts that are not really about you. Then run each one against the engines you care about and record the answer given, which sources were cited, and whether you appeared. Keep the prompt list fixed from that point on, because the entire value is running it again later and seeing what changed. Rewording the prompts each time makes the results incomparable, which is the most common way this measurement quietly becomes useless.

Answer in the first fifty words, then prove it

Retrieval pulls passages, not pages. Put the direct answer immediately under the heading that asks the question, in two or three sentences that stand alone without the surrounding article, then spend the rest of the section proving it. A section that opens with background, defines its terms for three paragraphs, and reaches the answer at the bottom gives a retriever nothing clean to lift. Test it the hard way: copy one section out of your page, paste it somewhere with no context, and ask whether it still answers the question and still identifies who said it. If it reads as a fragment, rewrite the opening sentence so it is self-contained. This also means repeating the subject instead of leaning on pronouns across sections, since a passage beginning with it depends on a sentence the model may never have retrieved alongside it.

Make every claim attributable

These systems favor sources that are specific and checkable, and so do the people who audit the answers afterwards. Every substantive claim on the page should carry a number, a date, a named source, or a described method. A sentence that states the sample size and the month the data was collected is quotable. A sentence that says studies show is not, and will lose to a page that names the study. If the data is yours, say how you collected it and when. If it came from elsewhere, link the original rather than a blog post about the original. Date the page visibly and update that date only when you actually update the content, because a page claiming freshness while carrying stale figures loses trust the moment someone checks one. Never invent a figure to fill a gap: a wrong number that gets quoted is far worse than no number. Quantify your own hedges too. Frequently, usually, and most teams are unquotable, where three of the five accounts we migrated last quarter is quotable, assuming it is true and you can show the working. Write the number you actually have rather than the one that sounds better.

Structure the page so a retriever can find the part it needs

Write headings as the question being asked, in the words someone would use, and keep one idea under each. Within a section, lead with the answer, then the evidence, then the caveat. Use short paragraphs, because a four hundred word block gets retrieved as one lump and mostly discarded. Put comparisons in tables with the compared attribute named in the header row, since a table is unusually easy to extract a single fact from. Write definitions as one sentence that begins with the term itself, so the extracted sentence works on its own. Where it fits, mark the content up with the relevant schema types for articles, products, and question and answer pages, which guarantees nothing but makes your structure explicit instead of inferred. Avoid burying the facts that differentiate you in prose that requires reading the whole page to assemble, because nothing will assemble it for you.

Earn the mentions these systems read

A large share of what an AI answer says about you comes from pages you do not control: review sites, comparison roundups, forum threads, community answers, documentation hosted elsewhere. Ask an engine to describe your product and you will usually see the sources it leaned on, and often they are third-party lists rather than anything of yours. So the work splits in two. On your own site, be the clearest available statement of fact about what you do. Off your site, make sure the pages that get cited are accurate and current: claim and correct your listings in the directories that rank for your category, get included in the roundups already appearing for your prompts, and answer the recurring questions in the communities where your buyers ask them. Correcting one stale third-party page that says you lack a feature you shipped a year ago is frequently the highest-leverage hour available.

Let the crawlers in and check what they see

None of this matters if the retrieval layer cannot read the page. Check your robots rules for the user agents these crawlers use and decide deliberately whether to allow each one, since blocking them is a legitimate choice but an accidental block is not. Then check rendering: fetch the page the way a simple crawler would, without JavaScript, and look at what comes back. If the answer you carefully placed in the first fifty words only appears after a client-side fetch, assume it does not exist. The same applies to content behind tabs, accordions that load on click, and anything gated by a consent banner. Confirm the pages you care about are indexed in conventional search as a baseline, because most of these systems lean on a conventional index. Fix crawlability before touching the writing, since no amount of writing helps a page nobody can read. Check a recently published page as well as an old one, because rendering behaviour frequently differs between templates built at different times by different people. Then fix the template rather than the page, or you will be doing this again next quarter.

Measure citations on a schedule and expect variance

Run your fixed prompt list monthly across the engines you care about, and record three things: whether you were cited, who else was, and what the answer said about you. The third matters most, since a citation attached to a wrong description of your product is a problem rather than a win. Expect noise. The same prompt can return different answers and different sources on consecutive runs, so treat a single result as an anecdote and look at the rate across runs instead. Watch the competitor set in those citation lists the way you once watched the top ten, because a name that starts appearing everywhere is doing something you can go and read. When you lose a citation you held, check whether your page changed, whether the cited competitor published something newer, and whether the answer format itself changed. Keep the interval fixed: this only becomes useful as a series.