For twenty years, being found meant ranking — earning a position on a page of ten blue links and letting the customer pick you. That contest is quietly being replaced. When someone asks ChatGPT, Perplexity or Google's SI mode for "a good florist that delivers same-day near Richmond", they do not get a list to browse. They get an answer, with one or two names in it. Being ranked was a traffic game. Being recommended is a trust game, and it is played by different rules.
The temptation is to treat this as another dark art with another acronym — generative engine optimisation, GEO — and wait for an agency to sell you the ritual. Resist that. What the answer engines reward is unusually legible, and most of it favours small, specific businesses over large, vague ones. The machines are trying to give a confident, checkable answer to a specific question. Your job is to be the easiest business in your category to be confident about.
What the machines are actually reading
An SI answer is assembled, not conjured. Behind the scenes the engine is retrieving and cross-checking sources: your own site, review platforms, directories, maps listings, press mentions, and the structured data your pages carry. When those sources agree — same name, same offer, same opening hours, same claims — the engine can repeat you without risk of embarrassment. When they disagree, it reaches for a competitor it can verify instead.
This is why specificity is the small operator's advantage. A model asked for "the best trainers" has a thousand safe answers. Asked for "a decaf subscription that is actually worth drinking" or "a two-chair salon in Leith that takes bookings by WhatsApp", it has very few — and a business that has said exactly that about itself, consistently, everywhere, is the natural completion of the sentence.
- Make one page the definitive answer to one question a customer actually asks — not a page that gestures at six services in general terms.
- Keep the boring facts identical everywhere: name, location, hours, delivery area, what you sell. Inconsistency reads as unreliability to a machine doing cross-checks.
- Carry structured data — product, price availability, reviews, business details — so engines do not have to guess what your pages mean.
- Accumulate real reviews and answer them. Third-party corroboration is the closest thing the answer engines have to references.
- Write in plain, factual sentences. Superlatives are unquotable; specifics are citations waiting to happen.
Write pages that can be quoted
Traditional SEO copy was written to be crawled and skimmed. Answer engines lift sentences, so the useful mental test changes: could a machine quote two sentences from this page and have them stand alone as a correct, complete answer? If the honest answer to "do you deliver on Sundays" is buried in paragraph five, the page ranks in theory and is unquotable in practice.
Answer first, evidence second
Put the direct answer in the opening lines — the yes, the number, the how-long, the price range — then spend the rest of the page earning it. Humans skim in the same order the machines extract, so nothing is lost.
One page, one question
A page that answers "do you do vegan cakes", "what is your delivery radius" and "can I hire you for weddings" at once answers none of them cleanly. Split them. Small pages that each settle one question are what retrieval systems are built to find.
The corpus you cannot fake
There is a hard edge to all this that keeps it honest: you cannot astroturf corpus-wide consistency. One suspiciously glowing page is easy to manufacture; agreement between your site, your reviews, your directory entries and what customers say about you in public is not. The businesses that surface in SI answers tend to be the ones that were already coherent — they simply made the coherence machine-readable.
It also means the fastest wins are usually corrections, not creations. An old phone number on a directory, a delivery policy that changed two years ago and was updated in only one place, a service you quietly stopped offering — each contradiction is a reason for an engine to hedge. An afternoon spent reconciling your public facts often does more than a month of new content.
Where your storefront does the work
A Phoxta storefront is built answer-shaped by default: product, availability, FAQ and review content live in the database and render as structured pages, so a fact corrected once — a delivery cut-off, an opening hour, a returns window — is corrected everywhere at the same moment. And there is a useful symmetry with the SI agent that answers your customers on web chat, SMS, WhatsApp and email: the questions people ask it are, almost word for word, the questions they ask the answer engines. The knowledge you load into one is the editorial brief for the other.
Search rankings measured how loudly you could compete for attention. SI search measures whether a cautious machine is willing to repeat you. Those are very different tests, and the second one favours the honest.
How to know it is working
You cannot install an analytics tag inside someone else's model, so the measurement is more manual and more truthful. Once a month, ask the major assistants the five questions a ready-to-buy customer would ask in your category, and record who gets named. Watch branded and direct traffic — people who arrive already knowing your name were often handed it by an answer. And listen in your own conversations for "the SI recommended you", which customers increasingly volunteer unprompted.
None of this requires a new discipline, an agency retainer or a rebuild. It requires being specific, being consistent, and running on a storefront that presents the truth in a form machines can verify. See what a running Phoxta business includes on the pricing page — the recommendable part is standard equipment.
