A home buyer does most of their shortlisting before they ever call a sales office, and a growing share of it now happens inside an AI answer rather than a listing page. An assistant can restate your registration number, approval status and configuration mix accurately, because those are defined facts with a public register behind them. It cannot reliably state your possession date or your all-in price, because both are claims that need a basis and the engine is quoting whatever it last read. Publish the durable half as text, and the fragile half with its basis attached, and you control most of what an answer says about your project.
The shortlist is made before the site visit
The old sequence was portal, enquiry form, sales call, site visit. The enquiry was where the buyer entered your world, and everything before it was somebody else's page. That sequence still exists, but a step has been inserted in front of it: the buyer asks an assistant which areas and which projects fit a budget, and arrives at the portal already holding a shortlist.
This matters more in real estate than in most categories because the buyer is not comparing products, they are comparing risk. A flat is the largest purchase most households make, and it is frequently bought before it exists. The questions that decide it are not about the flat at all. They are about whether the builder finishes.
So the answer a buyer reads is doing something a brochure never did. It is assembling a view of your reliability from sources you do not control, and presenting it as a summary, in one paragraph, with no room for the caveats your marketing copy relies on.
What does a home buyer actually ask before booking?
One line typed into an assistant is not one question. A prompt such as best areas to buy an apartment for families fans out into location and daily commute, total cost beyond the per-square-foot price, approvals and title, the builder's delivery record, the possession date and what happens if it slips, and carpet area against super built-up. The engine builds its answer from whichever of those it can verify, and quietly drops the ones it cannot.
Two of the six are about you rather than the property. A buyer asking how to check a builder track record before booking a flat is not researching a project, they are auditing a company, and the sources that answer them are local news, resident forums and the public register rather than your site.
The last of the six is the one that quietly decides comparisons. Carpet area and super built-up are different numbers for the same flat. An assistant handed one of each from two projects is comparing nothing, and a careful one will decline to compare at all.
The half an assistant answers well
Some project facts are stable, defined and publicly registered. The project registration number. The approval status. The declared completion date on the register. Configurations and carpet area. The number of towers and units. None of these are opinions, and none of them move between one crawl and the next.
An engine handles them well when it can read them. It reconciles your project pages against the public register, listing portals, maps and locality data, and whatever local coverage exists. Where those sources agree, the answer is confident and it cites somebody.
The question is whether it cites you. A configuration table published as text is readable. The same table inside a brochure image is not, and whoever retyped it into a portal listing earns the citation instead. Your project is then described second-hand, on a page you do not control, next to projects you did not choose to sit beside.
Why does a possession date go wrong in an answer?
Because a possession date is a forecast, and the engine cannot tell a forecast from a fact. It reads a date on your page, a different date on the register and a third in a news report about a delay, and it has to choose. What it usually does is hedge, and a hedged possession date reads to a buyer as a warning even when your project is on schedule.
The second failure is quieter. A date with nothing behind it cannot be checked at all. No stage of construction, no registered completion date beside it, no record of what you handed over before. A careful engine drops it, and dropped reads as though the question has no answer, which is worse for you than an honest date that is later than you would like.
The fix is not a more confident date. It is the same date with its basis in the same line: the registered completion date, the current stage, and your record of handovers with their actual dates. A forecast that shows its working survives being quoted. A bare forecast does not.
Total cost is the line that survives between crawls
If the possession date is the fragile half, the cost structure is the durable half, and it is under-published almost everywhere. Base price, floor rise, parking, club charges and registration usually live on three different pages, or in none of them, so an assistant assembling a cost picture stitches it together and will sometimes stitch it wrong.
Stated together, as text, they become the most quotable block on your project page. Cost is comparable by construction, which is exactly what a comparison answer needs, and the structure is still true next month even when the base price moves.
There is a second reason to publish it in full. A buyer who is convinced and cannot see how to begin is a lost booking. The site-visit path belongs on the same page, because an answer that recommends you should also be able to say what happens next, and a callback form is not a next step an assistant can complete.
What to publish, and what to read in the answer
Five blocks carry a project page: what is actually for sale, with carpet area and stage; the all-in cost with every charge named; approvals and registration stated plainly; what you have handed over before, with dates; and a real way to visit. Four of the five are facts you already hold. The fifth is usually a form.
Then read the answers. Run the prompts a buyer types, on the engines you care about, and record which rung you reach: absent, mentioned, cited, compared or recommended. The useful finding is rarely the rung itself. It is which source the answer leaned on, because that names the page that is speaking for you.
The pattern worth watching is a project that is named but never cited. It means engines know you exist and do not trust your page enough to quote it, which is almost always a specification sitting inside an image.