If your domain is new, the honest read is that you cannot win the terms you want for a long time. The sites holding them have been accumulating authority for years and have people whose job is keeping them there.
The useful question is what moves faster than domain authority. The answer turns out to be the surfaces answer engines pull from — and everyone worked that out at once, which you can watch happen in the search data:
| Keyword (US) | Monthly volume | Year over year |
|---|---|---|
| generative engine optimization | 12,000/mo | +161% |
| AI search optimization | 4,800/mo | +152% |
| AI visibility | 3,300/mo | +258% |
| AI overview optimization | 900/mo | +625% |
Those are Ahrefs' mid-2026 US figures, and the same analysis argues that generative engine optimization — GEO — is a subset of SEO rather than a new discipline. That reading is correct and it matters for everything below: there is no separate playbook, no file Google requires, no markup that generative results ask for and ordinary search does not. What changes is which surfaces carry the answer, not what good work looks like.
This is written from the position of running a new domain, not from having finished. Take it as a plan with reasoning attached rather than a report of results.
Why citation is the faster game
Ranking rewards accumulated trust in a domain. That is a stock, you build it slowly, and a new site has none.
Citation works differently. When an engine assembles an answer it is looking for passages that address the question directly and can be attributed to something. Authority still matters, but so do things a new site can control immediately: whether your claim is stated plainly enough to lift, whether it appears anywhere the engine already trusts, and whether it exists anywhere else at all.
| Dimension | Ranking | Citation |
|---|---|---|
| Dominant input | Accumulated domain authority | Whether a passage answers cleanly and is attributable |
| Time to move it | Months to years | Weeks |
| Where the work happens | Your own domain | Mostly surfaces you do not own |
| What a new domain controls | Close to nothing | Most of it |
| What blocks you | Everyone who started earlier | Nothing, if the information is genuinely yours |
The last one is the leverage. If the information exists in ten places, an engine has no reason to name any of them. If it exists in one, naming it is the only option.
Most of the work is not on your site
This is the uncomfortable part for anyone who would rather be writing.
Answer engines lean heavily on third-party surfaces — review platforms, comparison listicles, community threads, directories. Those surfaces have the authority you lack, and unlike your domain age you can appear on them in weeks. Getting listed where your category gets compared, and being present in the places your buyers actually ask questions, does more in the first months than another post will.
The order that follows from this is uncomfortable but clear: the off-site work outranks the on-site work early, and it costs time rather than engineering. If you are choosing between publishing a fifth article and claiming your profiles on the surfaces that already rank, claim the profiles.
Write only what a model cannot already say
The rule that survives contact with this: if a competent model can answer the question without you, writing the answer is volunteering to be summarised without attribution.
That rules out the explainer content most new blogs start with. Nobody needs the thousand-and-first description of what an AI agent is, least of all a model that has read the other thousand.
What it leaves is narrower and better. Things you measured. Methods you used and can describe precisely. Decisions you made with the reasoning intact, including the ones that turned out badly. Constraints you hit that are not documented anywhere because nobody else built the thing.
Original data is the strongest version of this, because a number that exists in one place must be attributed to that place. It is also the hardest, since it requires having enough of something to count. It is worth gating honestly: numbers computed from too small a sample are noise dressed as research, and publishing them is worse than publishing nothing, because the whole point was to be citable rather than dismissible.
The mechanics that matter on the page
Less than people think, but not nothing. Four things, in the order they pay:
- Put the answer first. The figure everyone quotes here — that something like 44% of citations come from the first 30% of a page — circulates with no methodology, sample size or date attached to it, so treat it as directional and do not repeat it as a fact. The underlying point stands without it: retrieval reads a page in order, and an argument that arrives in the sixth paragraph has wasted the position that was doing the work. Every post on this site puts its thesis in a panel above the body, and the argument alone is reason enough.
- Make passages self-contained. Retrieval operates on passages, not documents, so a section that only makes sense after the three before it is a section that cannot be lifted.
- Mark up what you actually have. Article metadata with a real modification date, and question markup only where the questions are genuinely rendered and genuinely answered. Marking up questions you did not answer is the kind of thing that stops working and then counts against you.
- Keep dates honest. Restamping a file changes nothing about the content, and the engines are not the audience being fooled — you are, about whether the post is still true.
You cannot measure any of this by default
The part most easily skipped: without the property verified in the search consoles, none of the above is measurable. You will not know which pages are indexed, what they surface for, or whether anything changed.
Two consoles, not one, and both are free.
Google Search Console covers Google Search, AI Overviews and AI Mode. Measure the classic results and the generative ones separately in it: overlap between them in your own data is an observation about your site, not a rule about how inclusion is decided.
Bing Webmaster Tools covers the Bing index, which OpenAI's search has drawn on. Nobody outside that arrangement knows its terms and it could change; while it holds, Bing's console is the closest thing to a free readout of one major assistant's retrieval, and that is worth an afternoon of verification.
And the cheapest honest measurement is manual: once a month, ask the assistants the ten questions your buyer would actually ask, and write down whether you appear. Ten minutes, no tooling, and it is the only thing that directly measures the outcome you are chasing.
What this adds up to
Expect little from your own domain for months. Spend that time on the surfaces that already have authority, publish only what could not have been written without you, and verify the properties so you can tell the difference between working and waiting.
Our own attempt at the third category is what we hit implementing a compliance requirement, which is the kind of thing that only exists because we did it. The free tools are the other half of the same idea: utility that cannot be summarised, because you have to run it.
Last reviewed 21 September 2026.