Nobody can find your launch: answer engine optimisation for new products

A launch can go perfectly and still be missing from the answer

You can ship on schedule, land the coverage, top your target community, and still be absent from where a growing share of buying research now begins: inside an assistant. Someone asks ChatGPT for the best tool to schedule shifts across a 40 person restaurant group and gets four named products, a paragraph each. You are not among them.

Nothing tells you this happened. A ranking drop shows up in Search Console, but an answer you were never included in produces no data: no impression, no click, no session. The buyer built a shortlist in ninety seconds, and you were not considered rather than rejected. This piece is the foundation for the rest of the AI visibility section: what you can influence before launch day.

Discovery and accuracy are two different problems

Almost everyone worrying about AI visibility is worrying about discovery: whether an assistant names you when a stranger asks about your category. Those unbranded questions come from buyers with no shortlist yet, which is why they are worth so much.

Accuracy is the second problem and gets far less attention. A prospect who has heard of you asks what you cost, and the assistant answers confidently from whatever it retrieved: a price you moved off a year ago, or a free tier you retired.

Discovery failures cost you consideration from people who might never have found you. Accuracy failures cost you deals with buyers who were trying to buy, and they compound, since a wrong fact repeated across three review sites becomes the consensus every assistant retrieves. Test the two separately, with unbranded questions and then branded ones. The twelve prompt test makes that repeatable; the AI visibility check scores what sits behind both.

Why this is a separate game from search

Being competent at SEO does not make you visible to answer engines, and being new does not disqualify you. The second half is the opening a launch gets.

The mechanism is worth understanding rather than taking on faith. Answering a category question involves retrieval: the assistant turns it into several searches, pulls candidate documents, and synthesises an answer from what they say. Those documents are mostly not yours. They are review sites, comparison articles, community threads, directories and competitors' pages that mention you, and your own site is a minority of that set.

This inverts a familiar assumption. In search, authority accumulated over years is the moat and a three month old site fights uphill. In retrieval, what matters more is whether independent sources describe you clearly enough to be picked up. Googlebot access still governs AI Overviews; what changes is the order of the work.

Start with access, because access is free

Open your robots.txt and read it as though a stranger wrote it. This is the cheapest work available and the commonest cause of total invisibility. Two patterns do most of the damage: a blanket disallow set during staging that reached production, and a block on AI crawlers added by somebody who read that AI firms were scraping.

The second deserves care, because training crawlers and retrieval crawlers are different things with different consequences. Blocking training reduces the chance your content enters a model's weights, which costs you almost nothing in visibility. Blocking retrieval makes citation impossible, because the agent fetching a page to quote in a live answer is not the one gathering training corpora, and obeys a different token.

The Google-Extended row repays a second reading. Google-Extended governs whether your content is used for Gemini training and grounding. It does not control whether you appear in AI Overviews or AI Mode, which ride on ordinary Googlebot access and normal indexing, exactly as a blue link does. Most articles state the opposite and teams act on it. Check your CDN and firewall rules too, since a block there defeats a permissive robots.txt.

Entity resolution comes before trust

An engine that cannot work out who you are will not name you, however good your content. Before authority enters the picture comes a mechanical problem: resolving a name like Kite or Atlas to one stable thing in the world.

New products fail this for mundane reasons. The company is called one thing on the website, another in the app store and a third on LinkedIn, each differing by an "Inc" or a space. Nothing asserts a canonical identity, and no sameAs array connects the site to those profiles. Worst is the namesake collision, where a dead product from 2011 shares your name and has fifteen years of documents behind it.

The fixes are dull and mostly one-off. Settle an exact trading name and use it identically everywhere, capitalisation included. Emit Organization schema with a stable @id and a sameAs array listing every profile you control, and claim the profiles a company of your kind ought to have. Where a namesake exists, disambiguate in prose: category, location, year founded.

Mentions do more work than links

Before launch, spend your effort on being mentioned rather than on being linked to. Muck Rack's Generative Pulse series puts earned media at 84 per cent of AI citations, and Ahrefs, studying around 75,000 brands, found that brand web mentions showed roughly three times the correlation with AI Overview visibility that backlinks did. That second figure is correlational: it does not show that mentions cause visibility, and brands that get mentioned differ from those that do not in other ways. It is still a strong hint about where the signal sits.

Links are slow to accumulate and gatekept by people with no reason to help an unknown company. Unlinked mentions count for retrieval and are far easier to earn: a sentence in a roundup, or a name in somebody else's comparison table.

In the eight weeks before launch, getting mentioned means work you can schedule. Get listed accurately in your category's directories and in the integration directory of every platform you connect to, since those pages get retrieved far more than their traffic suggests. Give the authors of pieces that already list your competitors a checkable paragraph about you. Do two or three podcast appearances where the show notes name the product in text. You are assembling independent documents that describe you consistently, the input retrieval runs on.

Write so you can be extracted

Retrieval does not read your page the way a prospect does. It pulls a passage of a few hundred words as evidence, so a page that answers in its fourth paragraph supplies a passage with no answer in it.

Put the direct answer in the first two sentences under every heading, then explain and qualify underneath. Write a definition sentence for every concept you own, in the form "X is Y that does Z", because that shape gets quoted verbatim more than any other. Put anything enumerable into a table: pricing tiers, plan limits, integrations. Tables survive extraction with their structure intact, while the same facts in prose lose their associations once cut. Keep FAQ answers self-contained, since one opening "It depends on your plan" is useless once separated.

The technical counterpart is rendering. Content that only exists after client-side JavaScript executes is invisible to most retrieval fetchers, which do not run a browser, so serve that text in the initial HTML.

Comparison and alternatives pages are the highest-yield format

Comparison questions are what buyers ask assistants most, and what your marketing site is least likely to answer. "X versus Y", "alternatives to Z", "what should I use instead of the thing I have outgrown": late-stage queries needing a document that compares things. Publish nothing and the comparison retrieved is a competitor's or a review site's, which will not write your side of it. Doing it well means naming real competitors, describing them fairly, and tabulating what buyers weigh: how to write comparison and alternatives pages covers the structure and the usual mistakes.

Freshness is an operating discipline

Roughly half of AI-cited pages were updated within the previous thirteen weeks, and visibility decays as content ages past a quarter. That is a property of the retrieval systems rather than your industry, and it holds even when your subject has not changed.

What it implies is a review cycle rather than a publishing calendar. Every page you care about gets a scheduled look each quarter, asking what is now wrong, what a competitor shipped, and what readers want that the page lacks.

Changing the date without changing the content fails on two counts. It is deceptive, and readers who notice stop trusting the page. It is also detectable, since retrieval systems compare content across crawls and a modified timestamp on an unmodified body is cheap to discount.

Structured data, with the caveats stated plainly

Emit structured data, but be clear about why. Google stopped showing FAQ rich results for most sites in 2023 and HowTo rich results are gone, so adding FAQPage markup in the hope of accordions is several years out of date.

The reason to emit it now is machine readability. Schema states what an entity is, what it costs and who published it, without an engine inferring it from prose. Organization with a stable @id and a full sameAs array does real work on entity resolution, and Product and Offer state pricing unambiguously. Keep it in sync with the page, since immaculate JSON-LD on an evasive body loses to a plain page with no markup.

An eight week pre-launch sequence

Order matters more than effort. Access comes before mentions, since mentions point at pages nothing can fetch, and identity is settled before outreach, since inconsistent naming manufactures the ambiguity you were removing.

Thirty days after launch, run the questions again and compare against the eight week baseline. That comparison is the only AI visibility measurement you fully control. The sequencing sits inside the wider 90 day pre-launch runbook, and AI visibility is one of the eight dimensions the launch readiness assessment scores.

What not to bother with

llms.txt. Adoption is low, no major engine has announced support for it, and Google has said publicly that it ignores the file. Writing one takes twenty minutes, so ship it as a cheap hedge if you like, but expect nothing from it. A guide presenting llms.txt as the answer to AI visibility is recommending the cheapest item on the list.

Spamming mentions. Because mentions carry weight, somebody will propose manufacturing them at volume: generated guest posts, listings you write yourself, forum accounts recommending you unprompted. Retrieval systems weight their sources, so a hundred mentions on sites nobody reads count for little against one paragraph your buyers actually read. An assistant reproduces the tone of its sources too, so astroturfed praise makes you read as a firm that astroturfs.

Prompt keyword stuffing. Repeating "best CRM for small business" forty times achieves nothing: retrieval matches on meaning, and the quoted passage must stand alone.

Chasing every new file format. Proposed standards will keep arriving, each promising to be how you talk to answer engines, and most go nowhere. The durable work has not changed: be fetchable, be identifiable, be described consistently by sources other than yourself, and answer the question on the page in a form that survives being lifted out of it.