A prospective client asks an assistant which agency they should consider for local search work in Kathmandu. They get a paragraph of advice & three named sources. They never see a results page, never scroll, never click ten blue links. If you are not in that paragraph, you were not in the consideration set at all.
This is a real shift in how buying research happens, & it changes what optimisation means. You are no longer competing for a position on a page. You are competing to be the source a model chooses when it constructs an answer.
How these systems actually select sources
Strip away the mystique & there are two mechanisms, & they behave differently.
Retrieval. When a model needs current information, it issues searches, retrieves pages, & constructs an answer from what it finds. This is closer to traditional search than people expect, because the retrieval step is usually powered by a conventional search index. If you do not rank for the underlying query, you are not in the retrieved set.
Training knowledge. What the model absorbed during training. You cannot influence this directly or quickly, but it is shaped by how frequently & consistently your organisation is described across the wider web.
The practical consequence is important: traditional search visibility remains the primary entry ticket to retrieval-based answers. Generative optimisation is not a replacement for search fundamentals. It is a layer on top of them.
If you are invisible in conventional search, you will be invisible in generative answers. Fix the foundation before optimising for the model.
Write in a form a model can lift
Once your page is retrieved, the question becomes whether it can be used. Models extract discrete, checkable statements. Content that buries a claim in a discursive paragraph is harder to use than content that states it plainly.
Concretely, this means:
- Answer the question directly & early. A section that opens with the answer in one or two sentences, then expands, is far more extractable than one that builds towards a conclusion.
- Use headings that are questions. They map cleanly to the queries being answered & they help a retrieval system locate the relevant passage.
- Make each section self-contained. A passage lifted out of context should still make sense. Avoid pronouns that depend on the previous section.
- State specifics. Numbers, timeframes, conditions & named steps are citable. Vague reassurance is not.
- Use lists & tables for genuinely structured information. They are unambiguous to parse.
None of this requires writing badly for humans. Clear, well-structured explanatory writing happens to be exactly what these systems handle best.
Be a consistent, resolvable entity
Models are trying to work out who you are. Inconsistency across the web makes that harder & reduces the confidence with which you are cited.
The practical work here overlaps heavily with conventional foundations. Use one consistent business name everywhere. Maintain accurate structured data describing your organisation, services & locations. Keep your profiles on major platforms current & aligned. Ensure your site clearly states what you do, where, & for whom, in plain text rather than only in imagery.
Named authorship matters more in this context than it used to. Content attributed to a real person with a stated role & a visible track record is easier to weigh than anonymous content. Build genuine author pages & link them consistently.
Mentions elsewhere carry weight
Training & retrieval both draw on the wider web, not only on your own site. What other people say about you influences how confidently a model describes you.
That makes third-party presence a generative optimisation activity, not only a PR one. Being described accurately in industry publications, directories, association listings, community discussions & comparison content all contribute. So does being quoted as a source, which is one of the more durable reasons to run a digital PR programme.
Consistency across those mentions matters as much as volume. Conflicting descriptions of what your business does produce hedged, unhelpful answers.
Do not block the crawlers you want
A practical & frequently overlooked point. Several AI systems use their own user agents, & your robots.txt may be blocking them, sometimes by default in a plugin configuration you never reviewed.
Decide deliberately. There are legitimate reasons a publisher might block model crawlers, but a service business that wants to be recommended almost certainly does not. Check your robots.txt, check any firewall or bot-management rules, & confirm your position matches your intention.
Measuring something that resists measurement
There is no console for this yet, so build a manual practice.
- Define a query set. Twenty to thirty questions a real prospect might ask an assistant about your category, your service & your market.
- Run them periodically across the assistants your audience uses, & record whether you appear, how you are described, & which of your pages is cited.
- Log inaccuracies. When a model describes you wrongly, trace the likely source. It is usually an outdated page, an old directory listing or an inconsistency you can fix.
- Watch referral traffic from assistant domains in your analytics. Volumes are small but the visitors tend to be unusually well qualified.
- Ask new enquiries how they found you. Assistant referrals frequently arrive with no trackable source at all.
What not to do
Do not create content designed only for machines. Thin question-&-answer pages generated at scale are recognisable, they perform poorly in conventional search, & conventional search is the retrieval layer. Do not attempt to manipulate models with hidden text or instructions embedded in pages; it does not work reliably & it is trivially detectable.
& do not abandon the channels that work today for a channel that is still forming. Generative visibility is a growing share of research, not yet the majority of it.
The honest summary is that this discipline rewards the same things good SEO always did, applied with more precision about clarity & consistency. See how we approach generative engine optimisation alongside technical SEO, or ask us to run a citation audit on your category.