AI Overview

For a long time a search result was ten blue links. You asked a question and got back a list of pages that might contain the answer, and finding it from there was your job. Searching for a document tool that suits a 45-person team meant opening eight tabs and lining up pricing tables by hand. Half of searching was rereading the same paragraph in slightly different words.

The gap widened as queries grew from single words into sentences. A list serves someone who types "collaboration tool" perfectly well, but someone asking how many teammates a free plan allows wanted one number, not a list. The answer often sat in the middle of some page as a single line, and the reader still had to go in and dig it out.

An AI overview moves that digging onto the results screen. The answer sentence goes above the list of links, with the sites it drew on shown underneath. It's the place where all the tuning work in answer engine optimization becomes visible, and marketing teams talk about it constantly because what changes here is traffic.

When a query arrives, the engine first gathers relevant documents, has a model read them and write a fresh answer, then places that answer at the top of the page with links to its sources. Four steps.

1. The question gets split apart. A single line about a document tool for a 45-person team becomes several smaller questions that can be answered separately — pricing tiers, seat limits, permission controls.
2. Documents get gathered for each small question. The engine pulls a handful of pages from its search index that look capable of answering. A page missing from the index never becomes a candidate at all. An excellent help center that isn't indexed is simply invisible here.
3. The gathered documents get read and an answer gets written. Text isn't copied over; content from several documents is combined into new sentences, and each sentence stays linked to the document it came from.
4. Sources get attached. Three or four cited sites appear beside or below the answer, and a reader can click through to the original. Whose names land in those cards is the most direct outcome this whole area of work aims at.

Step 3 is what makes this area behave unlike ranking. A results list returns roughly the same order for the same query, but an overview is written fresh every time, so asking the same question two days later produces different wording and a different mix of cited sites. Being in the sources today guarantees nothing about tomorrow.

That makes step 2 the part worth working on. A document that never enters the candidate pool can't be cited no matter how well it's written. The site has to be properly indexed, and what it covers has to be easy to recognize. Putting an llms.txt file at the site root to list where the important documents live, or marking up which block is a question and which is an answer with structured data, is the work of widening that path.

The term comes up most often when rankings held steady but visits dropped. Three situations pull it into a conversation.

1. Reading traffic — impressions climbing while clicks fall is the signature. More people are getting their answer on the screen and clicking nothing, which is zero-click search, and a rank report alone shows no problem at all.
2. Rewriting documents — two standards separate documents that get cited from documents that don't. Answer extractability asks whether a paragraph still works as an answer when lifted out on its own, and citability asks whether it's organized to be quoted, with dates attached to numbers and evidence attached to claims. A paragraph opening with "the second of these" can't be lifted out, because it needs the paragraph before it.
3. Reporting results — the unit of measurement shifts from rank position to whether you were cited. Keep a list of questions customers actually type, run them on a schedule, and count how often your name appears in the source cards. Recording four out of thirty each month shows whether the edits did anything. Without that record a team keeps rewriting documents on instinct and never learns which change moved something.

"We're still ranked first and visits are down 30 percent. The overview answers the whole question."

The drop came from a layout change rather than a ranking change.

"This paragraph needs the one above it, so it never gets cited. Put the answer in the first line and rewrite."

The fix belongs at the paragraph level, not the document level.

"Last month we showed up as a source on four of the thirty questions."

Citation counts are being tracked as a performance number.

At Pagedeck, Ravi owns that number. He keeps a list of thirty questions customers search before they buy, runs them himself every two weeks, and records in a table whether Pagedeck appears in the source cards under each answer.

The first measurement came back at one out of thirty. There was no shortage of help documentation, but most of it was written to follow the screen step by step, so no paragraph stood on its own as an answer. Ravi and Mia picked twelve documents and rewrote them so each paragraph opened with its answer. Two months later the same list came back at six. Nothing new was published; the same information was reordered inside each paragraph.

The most common reading is that overviews end search engine optimization. The link list got pushed down the page, so competing for position looks pointless — which is plausible if you only look at the layout. But the candidate documents a model reads come out of that same search index. A page that isn't indexed, or sits far down in it, never gets read and never gets cited. Drop search optimization and within a few months your name disappears from both the list and the overview.

The other common reading is that a citation brings visits with it. Appearing in a source card feels like it should convert into clicks, but most readers take the answer and close the tab. A citation is closer to an impression than to a visit. Judging this channel by visits alone leads teams to conclude nothing is working and stop editing, while brand recognition is in fact building. Count citations and visits as two separate numbers. The first says whether the content is being trusted, the second says whether anyone needed more than the answer.