Answer Engine Optimization

For a long time a search query returned the same thing no matter what you typed: ten blue links. That made the marketing job easy to describe. Get your page into those ten, and as close to the top as you can. Search engine optimization (SEO) is the craft built around that goal. You put the words people actually type into your titles and body copy, you get other sites to link to your page, you make sure the page loads quickly enough that nobody gives up waiting. Checking your work was simple too. You typed the query, looked at the search engine results page (SERP), and counted down to your own listing.

Then something that wasn't a link started appearing above the list. Search engines began reading several pages and writing their own summary paragraph at the top of the results. At the same time people stopped going to a search box at all and started asking a chatbot instead. Someone who gets a good answer that way reads it and closes the window. Your article can supply every fact in that paragraph and still receive no visit. The old chain — higher rank, more clicks, more signups — breaks in the middle.

At Pagedeck the break showed up in a routine traffic review. Ravi was pulling quarterly numbers and found an article titled something like "how to write up meeting notes" that had held third place in search for six straight months. Over those same six months, visits to it had fallen from about 4,000 a month to 1,600. The rank hadn't moved. Search volume for the query hadn't dropped either. When Ravi ran the search himself, there was a summary paragraph sitting above the list, clearly assembled from material in the Pagedeck article, with three sources listed underneath it. None of them was Pagedeck. The article was being used as raw material and credited to nobody. Nothing in the old dashboard explained the drop, because the old dashboard only tracked position and clicks, and position had not changed.

Answer engine optimization (AEO) is the name for the work aimed at that gap. The target moves from rank to citation. What you are competing for is no longer one of ten slots in a list but a place in the small set of sources an answer is built from, and that changes which parts of a page matter. How many times a keyword appears matters less. Whether a single paragraph, lifted out of your page on its own, still answers the question matters more. So does whether the claim in it can be checked by whoever repeats it.

AEO is less a single technique than four strands tied under one name: the neighbouring concept that sets its boundary, the mechanisms you actually touch, the criteria that decide whether a document is good enough, and the surfaces where results become visible.

The neighbouring concept marks the edge of the work. Generative engine optimization (GEO) describes tuning content for engines that compose answers rather than list pages, and in practice the two names get used interchangeably in the same meeting. If you insist on a distinction, AEO emphasises the outcome — a question gets answered — while GEO emphasises the mechanism, that a model generates the answer. Neither name points at different work. Whichever one your team picks, the files you edit and the sentences you rewrite are identical, so it is worth choosing one and moving on rather than arguing about it. Vendors tend to prefer whichever term their product is sold under, which is why the same conversation can contain both.

The mechanisms are about opening a path for a model to read your site properly.

1. llms.txt — a plain text file at the root of your site that states what the site covers and lists the documents worth reading first. It works as a signpost. On a site with 300 help articles, a model without that file picks up a handful of pages more or less arbitrarily; with it, the model starts from the ten pages you put at the top of the list.
2. Structured data — machine-readable markup that labels which block on a page is a question and which is the answer, which line is a price and which is a publication date. To a reader the paragraph looks unchanged. To a parser it now carries name tags, which makes the right piece much easier to lift out. It also removes guesswork: a date written as "last spring" is ambiguous to a machine, while the same date in markup is not.

The criteria separate a document that gets cited from one that doesn't.

1. Answer extractability — whether a single paragraph still reads as an answer once you take it out of its surroundings. A paragraph that opens with "the second of the three points above applies here" needs the preceding paragraph to make sense, so it cannot travel. A paragraph that opens with "write up meeting notes the same day the meeting ends" is already an answer.
2. Citability — whether the document is put together in a way that makes it worth naming as a source. Numbers carry the date they were measured, claims carry evidence, the author is identifiable. Given two documents containing the same fact, a model tends to cite the checkable one. "Collaboration tools keep growing" loses to "as of 2025, 38 percent of Korean companies had adopted a collaboration tool," because the second version tells anyone quoting it exactly what to check if the claim turns out to be wrong.

The surfaces are where the work becomes visible. An AI overview is the summarised answer block at the top of a results page, with a short list of the sites it drew on underneath. Getting named in that list is the most direct result AEO produces. It looks similar to a featured snippet and behaves differently. A featured snippet is a passage cut out of one page, so there is exactly one source. An AI overview is new prose written after reading several pages, so there are several sources and the combination shifts from one query to the next. That instability is worth expecting. Appearing once is not a position you hold; it is a result you have to keep earning as the set of pages being read changes.

The four strands run in order and feed each other. Mechanisms open the path, criteria drive the rewriting, surfaces tell you whether the rewriting landed, and what you learn there shapes the next document. Teams that set up the mechanisms and leave their documents untouched get read without getting cited. Teams that rewrite everything but never check the surfaces have no way of knowing which edit worked, so the next article is written on instinct again.

"This piece is still ranked third and traffic is down by half. I don't think we're in the AI overview — can we look at it as an AEO problem?"

Rank hasn't moved, so the cause sits somewhere other than rank. The proposal is to rewrite the document into a shape that gets cited, rather than to add more keywords.

"Don't bury the conclusion at the bottom. Put it in the first paragraph as one line. Right now you have to read the whole thing to get the answer."

A request to restructure the piece so that one paragraph, on its own, answers the question.

"Let's ship llms.txt first and check in two weeks whether we're getting cited."

Open the reading path before doing anything expensive, and don't try to measure the effect the next morning. It is also a sequencing argument: do the cheap mechanical work now, push the rewriting to the following sprint.

At Pagedeck this work passes between Ravi and Mia. Ravi brings a list of ten articles that have lost traffic, and Mia picks out the ones that describe the product rather than the category. Their first pass shortens nothing and cuts a great deal. An article that spent three paragraphs on background before reaching the point gets the answer moved to the opening line and the background pushed down. An article covering five different questions gets split, with a subheading per question, since a reader arriving from an answer usually came for exactly one of them. Phrases like "as explained above" get replaced by a sentence that restates the thing in place, because a paragraph has to stand without its neighbours before anything can lift it out.

Verification is slow. A revised article usually takes several weeks to show up in answers, and asking the same question twice produces slightly different responses. So Ravi doesn't search once the next day and draw a conclusion. He keeps a fixed list of twenty questions, asks them every two weeks, and records how many of those answers list Pagedeck among their sources. Three out of twenty becoming six out of twenty reads as progress, even though no single answer proves anything by itself. Two months of edits with the count unchanged sends him back to check whether the twenty questions were the right ones before blaming the edits.

The most common mistake is treating this as a replacement for SEO. A new name for a piece of work reads like an obituary for the old one, and watching traffic fall makes that reading feel confirmed. But the documents a model reads when it composes an answer are the documents a search engine surfaced first. A page that search can't find doesn't get cited either. Teams that swap one for the other stop building links and tidying site structure, spend six months polishing sentences, and lose rank and citations together — as rank slips, they drop out of the candidate pool the model reads from in the first place.

The other frequent mistake is hearing AEO as an instruction to write shorter. Models summarise, so brevity sounds like an advantage. What actually decides the outcome is structure rather than length. A long article with a subheading per question and paragraphs that each stand alone is easy to quote. A short article where every paragraph depends on the one before it offers nothing to lift. Teams that only cut tend to cut evidence and numbers first, which thins out the credibility their content marketing spent years building, and what remains is an article that is neither quotable nor worth reading. The thing to change is order, not word count. Lead with the answer and put the reasoning behind it, and the piece can stay exactly as long while its first paragraph works on its own.