One answer engine quoted us; the other could not say what we do

The honest way to find out whether answer engines know about your site is to ask them, which is less obvious than it sounds. Crawl logs tell you a bot arrived. Search Console tells you a page is indexed. Neither tells you whether a model composing an answer in your category can say anything about you, which is the only question that matters for GEO.

So a day after the migration, with the crawl logs showing heavy traffic from every major agent, we asked two answer engines the same thing and got results far enough apart to be useful.

What Perplexity said

Asked what the platform is, it returned a description drawn almost verbatim from the new homepage’s about section — the applications, the kind of business they are for, the fact that the same firm staffs the engineering teams.

Then the harder test: a commercial question, phrased as a buyer would phrase it, naming no brands. Which companies offer out-of-home advertising management software for billboard operators in India. It returned a list of about ten vendors and included ours, with accurate operational detail — inventory across billboards, unipoles, gantries and digital screens, availability, proposals, geotagged proof of display, invoicing.

Every one of those specifics exists on a product page. None of them is a claim about being leading, trusted or end-to-end.

What ChatGPT said

With web search enabled, on the same day, it described a development, staff augmentation and digital marketing agency, and listed products from the previous era of the business. Its sources were mostly Google Play listings and pre-cutover marketing copy.

Then, about the platform we had just spent weeks writing pages for, it said in terms that it did not have enough public evidence to identify it correctly.

Same site. Same day. Same robots policy, which explicitly permits both. The difference was not access.

The gap was specificity, in both directions

Two things were happening at once, and it is worth separating them.

The first is index freshness. ChatGPT was reaching for what its search layer already held, and that was the old site — which was still being served for several legacy URLs because our redirects were returning 404 instead of 301. That is a plumbing failure and we fixed it, and it will resolve.

The second is the more interesting one. Perplexity had crawled the new pages and found passages it could lift and attribute. The answer it produced about the product was essentially a compression of sentences we had written. There was nothing to invent because there was something to quote.

Where a page said only that a product was comprehensive and scalable, no engine produced anything, because there is nothing in that sentence to carry into an answer. A model cannot cite an adjective.

What that implies for how you write

A search engine ranks a page. An answer engine lifts a passage. Those reward different things, and the second is a considerably stricter test.

A passage that gets quoted tends to have four properties. It answers a question somebody would actually type, rather than a topic somebody would browse. It contains a specific that could be checked — a mechanism, a number, a named failure mode. It stands up on its own, without the paragraph before it, because it will be extracted without that paragraph. And it is attributable, because it says something that is true of you rather than of the category.

The fourth is the one most marketing copy fails. A sentence that would be equally true of every competitor cannot be attributed to you, so there is no reason to cite you for it.

Limitations get quoted

The counter-intuitive finding, and the one we have leaned on hardest since: sections that say what something does not do get picked up disproportionately. Every product page here now has one, and every industry page has a passage about when the software is the wrong answer.

That is not a trick. It works because a qualifying statement is unusually safe for a model to repeat — it is falsifiable, it is specific, and it does not read as a claim that needs balancing. It is also, separately, the thing a buyer most wants to know and least expects to be told.

Training and retrieval are different pipes

One distinction worth holding onto, because it decides what is worth your effort.

GPTBot, ClaudeBot, CCBot and the rest are training crawlers. What they take may or may not enter a future model, on no timeline you can influence, and no model already trained can ever learn that your site changed. There is nothing to submit and nothing to expedite.

OAI-SearchBot, PerplexityBot and the search layers behind the assistants are retrieval crawlers. They fetch at or near query time. That is the pipe through which a new site can appear in an answer this month, and it is the one worth optimising for — which in practice means being crawlable as plain HTML, being fresh, and being specific.

In our first days after cutover the training crawlers were the loudest by an order of magnitude — several hundred requests each from ClaudeBot and GPTBot — while OAI-SearchBot, the one that could actually put us in an answer, had made ten. It is a useful corrective to how those numbers feel when you first read them in a log.

The test to run

Do not infer this from your analytics. Once a month, ask three or four answer engines the question a buyer would ask in your category, without naming yourself, and see whether you appear and whether what they say is right. It takes ten minutes and it is the only direct measurement of the thing you are actually trying to affect.

We found, doing exactly that, that one engine could describe our product accurately and another could not identify it at all — and that the second was reading a version of our company that had not existed for a fortnight.