Answer Engine Optimization: What Google Still Requires

Answer engine optimization is the work of getting cited when an AI writes an answer. Google still treats that work as ordinary search work, so the new name does not give your site a new set of rules.

Cited in this note

40%

Google’s guide says the work is still the same old search work.¹ The page has to be in Google’s list of pages, and it has to be able to show a snippet, which is the short piece of text Google puts under a result. Google says it does not use llms.txt, which is a plain file some people add to a site to tell AI tools what to read. It also says there is no special way you must cut up your text. And no extra hidden labels in the page code are needed for these features. We turn that into a short list of things you do not get. You do not get a magic file that makes you the answer. You do not get to hide the real page behind a block of text a person cannot read. What you do is plain. You publish a page Google can list, a page a snippet can lift from, and a page a customer can finish.

The visibility guide is the matching check for whether the site can be cited at all. The ChatGPT note covers crawler files, which are a different switch from this one. The GEO comparison puts the research and the click study side by side, if you want to see both.

What the research measured

A 2024 paper on generative engine optimization reports visibility gains of up to 40%.² Methods that add citations, quotes, and facts with numbers were linked to gains of more than 40%. On Perplexity, which is an AI search tool, the gain reached 37%. A score for how much of the answer came from the page rose 41%, and a score for how useful the answer felt rose 28%. The test used ten thousand questions. Those results come from pages written to be easier for an AI answer to lift from. They do not promise that your faucet page will move by a set amount because you added a quote.

When we sit with an owner, we take one thing from the paper. It is the kind of sentence that got the lift. A citation points at a source. A quote is a real person saying something you can check. A number with a clear group behind it is a fact the answer can carry. A page that says “we are the best hardware store in town” has none of those. A better page says how a cartridge is measured, quotes the maker’s note on fit, and tells the reader which handles in the aisle match it. The gain is tied to that kind of writing. It is not tied to the phrase “answer engine” in the footer.

Pew, in July 2025, found that people clicked a link 8% of the time when an AI summary showed up.³ When no summary showed up, they clicked 15% of the time. A citation that nobody clicks is still a citation. But you should write the page so the click is worth taking.

Owners feel that drop at the till, even if they never see the study. If the summary already told the person the cartridge size, they may not need you for that fact. They still need you for the part in stock and the person who will show them the wrench. They also need an honest line about what you will not do on a Sunday. Put that on the page. The summary can borrow the size, but the visit is for the counter.

A worked example

Walk into a hardware store on your street. This is an example we talk through, and it is not a result. One person works the counter and knows which aisle has what. People already ask a question like this, out loud and in search: can someone cut a key and tell me which cartridge fits my dripping faucet?

The old page says the store sells hardware, paint, and keys. It has a phone number and a cheerful line about friendly service. Google barely lists it, and a snippet would have nothing sharp to lift. An AI asked about that faucet will cite a page that states the size, the common mix-up, and what to bring to the counter. The owner’s page can be that page. But it will not get there by adding a new acronym to the title.

What the counter lead knows is the raw material. He already tells the person holding a bag of old parts that the stem length matters more than the brand on the handle. He also says a photo of the old cartridge saves a second trip. That line belongs at the top, in words a summary can quote and a person can use. Under it, the owner writes what the store really does. They cut keys while you wait on the days the counter lead is in. They match cartridges if you bring the old one or a clear photo. They do not fix pipes inside the wall. That last line is useful because it keeps the answer honest. It also keeps a click from turning into a sad visit.

Each person has a clear job. The owner owns the claims and will not let the page say they repair what they only sell. The counter lead checks the draft against what he says at the counter, because a polished sentence that does not match him will annoy the next customer. Whoever updates the site, even if that is a nephew, has one job. The page must stay listed, the first sentences must answer the question, and the source of any number must be easy to see. If a number is about their own stock, the page says it is theirs. If it comes from a maker’s sheet, the page names the sheet. That is the same move as the paper’s citations and quotes, done at the size of one shop.

An AI answer might then say the stem length matters and send the person to the owner. Pew’s finding is why the page cannot stop at the fact the summary will take. The summary can explain the stem, but the page has to make the trip better. Bring the old cartridge. The counter lead is the one who matches it. The store does not open the wall. Those lines are the reason to click, and they also keep the citation from turning into a complaint.

How we’d run it

We would not start with a tool that “does answer engine work.” We would start with the page a customer already lands on, and we would ask whether a stranger could act on it.

First, we check that the page is in Google’s list and can show a snippet. If it is blocked, thin, or tucked behind a form, we stop there. Google’s guide is blunt that these features still need an ordinary page. No special file can take the place of one.

Second, we write the question in the customer’s own words at the top, and we answer it in the next sentences. That might be key cutting, a cartridge match, or what to bring. A heading that repeats the question helps a person scan the page, and it gives an answer something clean to lift. This is plain structure and not a special ritual. Google says it has no special rule for how you cut up the text.

Third, we add the kind of material the 2024 paper linked to the gains. That means a real citation, a real quote, and a real number. For the owner, that might be the maker’s fit note, a sentence that really came from the counter lead, and a plain list of the sizes they stock. We do not invent a study. We also do not borrow a number we could not defend if a customer read it back to us.

Fourth, we add the reason to click, knowing a summary may already have stated the fact. That means the hours in words, what you will not do, who does the matching, and what to carry in. Pew’s July 2025 read is the background here. People clicked 8% of the time when a summary showed up, against 15% when it did not. So we write for the person who still has a reason to come.

Fifth, we skip the extras Google says it does not need. We do not wait on llms.txt. We do not add hidden labels in the page code that we cannot explain. We do not chop the page into scraps that only a machine could love. The person in the aisle is the reader who can hurt you.

Sixth, we read the page out loud at the counter. The counter lead fixes anything he would not say, and the owner deletes any boast. Then we leave the page up long enough to see whether the questions at the register start to match the questions on the page. At this size, that is the feedback that counts. A visibility score from a vendor is nice to have, but the counter is not optional.

Common mistakes

The first mistake is treating the new name as a new channel with its own file. Google says this is still ordinary search work, and it says it does not use llms.txt. A shop that spends the week on a special file and leaves the faucet page vague has done the opposite of what the guide says.

The second mistake is a page that nobody can lift from. It has no clear answer in the first lines and no heading a person can scan. Its only real fact is a phone number. Being able to show a snippet is a rule and not a matter of taste. If a snippet would be empty, an AI answer will be empty too.

The third mistake is fake precision. The paper’s gains were up to 40%, and above 40% where citations, quotes, and numbers were added. Those gains describe methods on a test of ten thousand questions. If you paste those percentages into the owner’s title, you tell a lie about her store. Use the method, but do not wear the result.

The fourth mistake is writing only for the summary. You polish the fact and forget the click. When a summary was there, Pew saw a click 8% of the time, against 15% without one. A perfect citation that sends nobody to the counter makes for a thinner business, even if a dashboard calls it visibility.

The fifth mistake is a page that says one thing while the counter does another. The page promises same-day key cutting. But the counter lead is off that day, and the backup will not touch the machine. The AI cites you and the customer shows up. Then you lose the one thing that mattered, which is being believed next time.

The sixth mistake is extra hidden labels as a superstition. The guide says no extra labels are needed for these features. Adding code you do not understand will not fix a page that never answered the question.

What we’d do

We would rewrite the one page people already need, in the counter lead’s own words, with the refusal included. We would ignore the pitch that this is a separate craft with a separate invoice. Our view is that the paper is useful and the label is not. The gains in a 2024 study are a reason to put citations, quotes, and numbers on the page when you truly have them. They are not a reason to talk differently from the way you talk at the counter.

Google still wants a listed page that a snippet can use. Pew says the click gets rarer when a summary shows up. So the page has to be worth the trip after the answer has already spoken. We would rather the owner be cited for a true line about stem length and then earn the visit. That beats chasing a file Google has said it does not read. If you do one thing after this, open the page that should win the question and see whether a stranger could follow it. That is the work. The acronym can wait in the hallway.