The New Referrer in Your Analytics
Open your analytics and look at the referrer list. Somewhere below Google you will find ChatGPT, Perplexity, Copilot, and Gemini sending real visitors to real pages. The volume is still small for most sites. The trajectory is not.
That traffic comes from citations. When an AI assistant answers a question, it increasingly names its sources and links them. Which sites get named is not random, and it is not the same contest as ranking in classic search. Answer engine optimization, AEO, is the discipline of winning that contest.
I have spent the past two years rebuilding client sites for this environment, including full prerender and structured data overhauls, and the pattern of what works is now clear enough to write down.
How Answer Engines Actually Choose Sources
An assistant answering a question does some combination of three things: it recalls training data, it runs a live search and reads the top results, and it consults an index it built by crawling. In the second and third cases, your site gets a few seconds of machine attention, and the machine is asking one question: can I lift a clear, attributable answer from this page?
That framing explains almost every tactic that follows. You are not optimizing for a ranking position. You are optimizing to be quotable.
Machines Do Not Run Your JavaScript
The single most common AEO failure I see is invisible content. Most AI crawlers do not execute JavaScript. If your site is a single page application that ships an empty div and hydrates it client side, GPTBot and PerplexityBot see a blank page wearing your meta tags.
I learned this the hard way on a chapter association site I rebuilt: the meta layer was perfect, and the served HTML contained nothing. We added server side prerendering so every route delivers real semantic content in the initial response. Citations and AI referrals followed. If you only do one technical thing after reading this piece, view source on your key pages and confirm the answer to the reader's question is present in the raw HTML.
Write Answers, Then Argue
Structure matters more for answer engines than it ever did for classic search. The format that gets lifted is: a heading that states the question, followed immediately by a two or three sentence direct answer, followed by the depth.
Most expert writing inverts this. It builds context for six paragraphs and lands the conclusion at the end, which is satisfying for a committed human reader and useless for a machine skimming for a liftable answer. You do not have to dumb anything down. Put the answer first, then earn it.
This is the same discipline I described in my breakdown of what ChatGPT did to search: synthesis queries are being answered in the interface, so the content that survives is the content that is either the best source for the synthesis or too specific to synthesize.
Be the Primary Source or Be Skipped
Answer engines are synthesis machines. They do not need a fourth blog post restating common knowledge, because common knowledge is what they are best at generating themselves. They need sources for things they cannot generate: original data, first hand experience, named expert judgment, and current facts.
Practically, that means:
- Publish numbers you own. A small original survey, your own campaign results, a measured before and after. Data with a named source is citation bait in the best sense.
- Attach a person to the claim. "Pierre Subeh, who ran the campaign, says" is liftable. "Experts agree" is not.
- Date your claims. Assistants care about freshness because their training data is stale. A clearly dated page about a current topic wins citations from an undated one.
- Take a position. Assistants asked for recommendations look for sources that actually recommend something.
- robots.txt that admits AI crawlers. GPTBot, PerplexityBot, ClaudeBot, and Google-Extended obey robots directives. Blocking them is a legitimate business choice, but make it a choice, not an accident.
- Structured data that matches the visible page. Article, Person, Organization, FAQPage, and Event schema give machines the entity graph. Keep it truthful; a mismatch between schema and visible copy reads as spam.
- An llms.txt file. An emerging convention: a plain text index at /llms.txt that tells language models what your site contains and where the canonical pages are. Cheap to add, and early adopters get disproportionate attention.
- Clean canonical signals. Assistants deduplicate aggressively. Syndicated copies without canonical attribution can outrank your original in an answer engine's source list.
- Fast, static, crawlable HTML. Every second of render time is a reason for a bot with a crawl budget to move on.
- Referral traffic from assistant domains, which most analytics tools now classify.
- Citation spot checks. Ask the major assistants the twenty questions your customers ask, monthly, and record which sources get named. This is tedious and worth it.
- Branded search lift. People often see your name in an AI answer and then Google you. Watch branded query volume alongside assistant referrals.
- AI assistants cite sources they can read and quote. Most cannot run JavaScript, so your answers must exist in raw HTML.
- Lead with the answer, then provide the depth. Question shaped headings with direct answers underneath get lifted.
- Original data, named experts, dated claims, and clear positions earn citations. Restated common knowledge does not.
- Do the boring infrastructure: permissive robots.txt for the crawlers you want, truthful schema, llms.txt, clean canonicals.
- Measure with assistant referral traffic, monthly citation spot checks, and branded search lift.
The Technical Checklist
The infrastructure work is unglamorous and decisive:
Measuring Something That Resists Measurement
There is no Search Console for ChatGPT yet. You can still measure:
I walk through the audit process in more detail in my AI Overview strategy piece, which pairs with this one.
Key Takeaways
Search is not dying. It is splitting. The brands that win the next five years will be visible in both halves.