A client of mine held position one for a commercial keyword for three straight years. Then, over about five months, their clicks on that term dropped by roughly a third while their ranking never moved. Nothing broke. The page was simply being read by a machine and summarized to the user before they ever reached it.
That is the whole story of generative engine optimization. The ranking is intact. The click is gone. What replaces the click is a mention, and mentions are earned by a different set of mechanics than blue links.
I have spent the last two years rebuilding how X Network approaches content for clients because of this shift. Here is what actually works, and what is mostly consultants selling fear.
What a generative engine is actually doing
Strip away the branding and every AI answer engine does four things in sequence.
It rewrites the user's messy question into several cleaner queries. It retrieves a candidate set of documents for each of those queries. It extracts the passages that most confidently answer the question. It composes an answer and attributes some of those passages to sources.
Every one of those four steps is a place you can win or lose. Most SEO advice only addresses step two, which is retrieval, and retrieval is the part you were already good at. The money is in steps one and three.
Step one is query fan-out, and you should optimize for the fan, not the seed
When someone types "best CRM for a small agency," the engine does not search that string. It generates half a dozen sub-queries: pricing, user limits, integrations, migration difficulty, whether it handles retainers. Then it assembles.
This means a page built around one keyword loses to a page built around a question cluster. I stopped briefing writers with a keyword and started briefing them with a list of eight to twelve sub-questions the buyer will actually think of. If your page answers ten of the twelve and your competitor answers four, you get cited in more of the composed answer even if they outrank you.
Build those clusters properly instead of guessing. I run seed terms through a keyword clustering tool and then manually add the questions that only someone who has sold the thing would know to ask. The manual additions are where the differentiation lives, because every competitor has access to the same tool output.
This is the same discipline as building topical authority maps, just applied at the page level rather than the site level.
The passage is the unit, not the page
The single biggest mental adjustment: engines do not cite pages. They cite passages. A 3,000 word masterpiece with one quotable paragraph performs worse than a 1,200 word piece with nine of them.
A quotable passage has four properties. It is self contained, meaning it makes sense if you rip it out of the article with no surrounding context. It is short, roughly 40 to 80 words. It contains a specific claim, not a hedge. And it sits directly under a heading that matches the question it answers.
Here is the test I make my team run. Take any paragraph from the draft, paste it into a blank document with no title, and ask whether a stranger would understand what it is about. If the paragraph starts with "This means that" or "As we mentioned above," it will never be cited, because it does not survive extraction.
Write definitions before you write arguments
Language models are disproportionately likely to quote a clean definition. If your page is about server side tagging, and somewhere in it there is a sentence that says "Server side tagging moves the tracking script from the user's browser to a server you control, so the data is collected once and distributed to platforms from there," that sentence gets pulled constantly.
I now require every substantial page to contain at least three definitional sentences of that shape. They cost 90 seconds to write and they are the highest yield sentences on the page.
Specificity is the ranking factor nobody lists
Generative systems are optimized to avoid saying vague things. When two sources both address a question and one contains a number, a threshold, a version, or a named constraint, the specific one wins the citation almost every time.
Compare these:
"Page speed is important for user experience and conversions."
"When Largest Contentful Paint crosses 2.5 seconds on mobile, the abandonment curve steepens noticeably, and in my experience the fix is almost always an unoptimized hero image rather than JavaScript."
The first is unciteable because it adds nothing to the model's existing knowledge. The second gives the engine something it can attribute to you. This is why generic AI written content performs so badly in AI answers, which is one of the funnier ironies of the current moment and something I dug into in why AI content still needs human writing.
The structural work that actually moves the needle
Four things, in the order I do them.
Answer first, then expand. Every H2 should be followed within two sentences by the direct answer. Then you elaborate. The classic magazine structure of building to a conclusion is actively hostile to extraction.
Make headings interrogative or declarative, never clever. "The Hidden Cost of Cheap Backlinks" is a great magazine headline and a terrible extraction anchor. "Cheap backlinks cost more than they return because reclamation takes six months" is worse prose and better performance. I compromise: clever titles, plain headings.
Ship real structured data. Not just Article schema. FAQPage where you genuinely have questions, HowTo where there are steps, Organization and Person markup so entity resolution knows who you are. I cover the deeper implementations in structured data beyond the basics, and it pairs directly with the entity work in entity SEO and knowledge panels.
Date and attribute everything. Visible publish dates, visible update dates, a real author with a real bio linked to a real profile. Retrieval systems weight freshness heavily on anything that could plausibly change, and unattributed content gets discounted.
What I stopped doing
I stopped chasing word count. Longer no longer means better and often means worse, because it dilutes passage density.
I stopped writing conclusions that summarize. Nobody reads them and no engine cites them. A conclusion should contain a new specific thing.
I stopped treating AI traffic as a separate strategy. It is not a channel. It is a distribution surface for the same asset, and the site level fundamentals in the technical SEO checklist still gate everything. If your pages render slowly or block crawlers, none of the passage craft matters.
I also stopped believing the tools. Rank trackers for AI answers are directionally useful and precisely wrong, because the answers vary by user, session, and phrasing. I track mentions manually across a fixed set of 30 prompts per client, run monthly, same account, same wording. Crude, but it is the only measurement I trust.
What to do Monday morning
Pick your five highest value commercial pages. For each one, write down the twelve questions a serious buyer would ask. Count how many the page currently answers with a self contained 40 to 80 word passage sitting under a matching heading. My bet is the number is two or three.
Get it to nine. Do not add a single new page until you have. In every account where we have run this, the mention rate moved before the traffic did, and the traffic followed within two quarters. The work is unglamorous and it is mostly rewriting things you already published, which is exactly why most of your competitors will not do it.