LLM Optimisation: How Your Store Shows Up in AI Answers

Someone looking for a quiet standing desk under €500 needs more than model names. They need comparable prices, noise levels and desktop dimensions from a source they can check. Shoppers ask those questions in ChatGPT, Perplexity and Google’s AI search too. How can a store make its answers useful there?

Key takeaways

• AI answers draw on three sources: training knowledge, live search results, and structured data that agents and shopping systems read directly.
• Google’s AI search can link to eligible product pages. No fixed rank or special markup guarantees a citation.
• Citable content – facts, tables, honest FAQs – gets cited; marketing fluff doesn’t.
• Honestly assessed, llms.txt is an unproven extra: costs nothing, does no harm – but Google declares it unnecessary and studies measure no effect.
• A full product feed supplies ChatGPT with current catalogue data. Display depends on processing and the query.
• Measurement has grown up: tools like Profound, Peec and Otterly automatically track mentions, citations and share of voice in AI answers.

How LLMs form recommendations

Three sources feed AI answers to buying-advice questions: the model’s world knowledge learned during training, live search results in web-enabled systems such as Perplexity, ChatGPT Search or Google’s AI search, and increasingly structured data – schema markup and product feeds – that agents and shopping systems read directly. You can invest in all three, just on different time horizons.

Training data works slowly, over months and new model versions – what matters most here is whether your content is present in the open web and citable whenever a model gets trained. Search integration, by contrast, works almost instantly: improve your classic ranking today, and it can show up in an AI tool’s very next live search. Structured data, finally, is the lever gaining weight fastest as agent usage grows – more on that in our article on agent-ready commerce.

What Google requires for AI Overviews and AI Mode

Google can link to several sources that address different parts of a question. For a standing desk, one may cover measured noise while another explains whether the motor supports a heavy desktop. That is a good reason to publish intelligible specifications.

According to Google Search Central, a page must be indexed and eligible for a search snippet. No special AI schema or extra file is required. Accurate availability, clear variants and verifiable specifications are more useful than promises of a fixed AI position.

An honest assessment includes the flip side: many AI answers end without a click. All the more important that the answer names and links your brand correctly – visibility is measurably shifting from clicks to mentions.

Creating citable content

LLMs love clear, fact-rich answers and avoid vague marketing language – comparison tables, specifications, honest pros and cons, and substantial FAQ blocks all get cited disproportionately often. The mechanism behind this is simple: a model composing an answer looks for extractable facts, not mood.

That has a direct consequence for product pages: thin, interchangeable content – identical descriptions across colour variants, for example – gets penalised by search engines and ignored even more thoroughly by language models, because it offers no extractable value. Our article on AI product copy at scale describes how to avoid this with copy generated at scale.

What works particularly reliably: structured comparison tables with clear column headings, numbered step-by-step instructions, FAQ blocks that open with a direct answer, and definitions that pin down a term precisely in one or two sentences before the deeper explanation follows. Plain prose paragraphs with no discernible structure, on the other hand, get cited less often, even when they’re just as factually correct – the model first has to laboriously distil the facts out of the text instead of lifting them directly.

From experience: a simple test shows whether a page is built to be citable: can you extract a correct, standalone answer to the heading from the first two or three sentences of a section? If not, the actual information sits buried too deep in the text for a model to reliably pick it up.

Establishing entity clarity

Brand, products and company need to be recognisable as clear, unambiguous entities so a model can correctly place them in context. In concrete terms, that means: consistent name spelling across every channel, clean Organization and Product schema as JSON-LD, and presence in sources models tend to trust – Wikipedia-adjacent sources, trade media, reputable review platforms.

In practice, that means writing your company name, brand name and product names identically on your own website, in marketplace listings, in press mentions and on review platforms – a company that shows up sometimes as “blicht GmbH”, sometimes as “Blicht” and sometimes as “blicht.com” makes it needlessly hard for models to map all three mentions to the same entity. A clean Organization schema with name, logo, contact details and social profiles on every page isn’t a nice-to-have here – it’s the machine-readable confirmation of that identity.

Machine readability and freshness signals

For ChatGPT search, you can allow OAI-SearchBot and separately block GPTBot for training. Also check whether the product data is readable in the HTML you serve. OpenAI’s crawler guide

Freshness signals have firmly joined the must-do list in 2026. For time-sensitive questions, models and their search backends visibly prefer well-maintained sources: a correct dateModified in your schema markup, lastmod values in the sitemap that match the real change dates, and visible last-updated dates on the page. According to one analysis of AI citations, around half come from content less than three months old. Especially relevant for stores: outdated prices, availability and statistics disqualify you – pages with numbers need a fixed refresh cadence.

Google also has a generative AI performance report in Search Console. For properties with access, it shows link impressions in AI Overviews and AI Mode. It does not count unlinked brand mentions or purchases, or report individual AI clicks or queries. A link impression answers a different question from a sample prompt in a chatbot.

llms.txt: an honest assessment

Hardly any GEO building block has been hyped like llms.txt – a curated markdown summary of your most important pages, meant as an entry point for AI systems. By now there is data, and it is sobering: Google’s 2026 guidance on AI search explicitly clarifies that llms.txt is not needed for AI Overviews or AI Mode and has neither a positive nor a negative effect. Otterly measured over 90 days that only around 0.1 per cent of all AI crawler requests fetch the file at all, and an analysis of roughly 300,000 domains found no measurable link between llms.txt and citation frequency in AI answers.

The honest line is therefore: costs nothing, does no harm, benefit unproven. We use llms.txt on blicht.com ourselves – as a maintained table of contents for machines, the half hour of effort is justifiable, and some AI vendors still recommend the convention for documentation sites. But it belongs at the end of the list, not the beginning: no substitute for structured data and citable content. If you only have one hour, put it into schema markup.

ChatGPT shopping: the product feed as a direct channel

OpenAI documents a full catalogue feed containing product identifiers, prices and stock. An eligible merchant first submits sample data for validation and then delivers regular full snapshots. OpenAI recommends updating at least daily. Uploading a price change does not establish that it already appears in a particular ChatGPT answer. Integration guide · Feed specification

After an update, check the exported data, the platform’s processing result and the price at purchase separately. For a store selling variants, a mismatched size or currency often tells you more than a count of ingested items. Agentic commerce guide

Collecting mentions, not just links

LLMs weigh where and how often a brand shows up in the context of a topic – not just who links to whom. PR, reviews and community presence therefore pay directly into AI visibility, even with no classic backlink at all. Customer reviews play an underrated role here, because they give models additional, credible mentions that extend beyond your own website.

The most common mistakes in LLM optimisation

Four patterns prevent visibility in AI answers in practice:

1. Search crawling accidentally blocked. Confusing training crawlers with search crawlers can affect discoverability. Record what each permission is meant to allow.
2. Only marketing language, no facts: product pages full of adjectives but without specifications give a model nothing to cite. Fix: add concrete facts and comparison data.
3. Inconsistent brand presentation: different spellings or contradictory information across channels prevent entity recognition. Fix: maintain name spelling and core facts centrally.
4. No monitoring: nobody regularly checks whether and how the brand shows up in AI answers – gaps only surface once revenue is noticeably missing. Fix: establish a fixed review cadence, via manual audit or a GEO tool.

Measuring: manual audit plus GEO tools

Before you optimise, measure your baseline: our free LLM Brand Check asks the major models in one minute what they know about your brand or domain – the basis for everything that follows. Building on that, a simple, repeatable audit process makes GEO visibility tangible:

1. Compile a list of your category’s twenty most important buying questions.
2. Ask these questions in ChatGPT, Perplexity and Google’s AI Mode every month and log whether and how your store gets mentioned.
3. Note the sources cited in the answers – they often reveal which of your own pages, or which third-party sources, actually get read.
4. Track referral traffic from AI sources separately in your own analytics, rather than letting it disappear into “other sources”.
5. Close any noticeable gaps (questions with no mention) with targeted, citable content.

What was pure manual work two years ago is now a mature tool category: specialised GEO trackers automatically measure whether and how a brand appears in AI answers – as a mention, as a cited source, and as share of voice against competitors.

From experience: for most stores, the pragmatic entry point is a combination: the monthly manual audit (one to two hours) plus an affordable monitoring tool. Investing in the bigger platforms only pays off once AI referral traffic in your own analytics is demonstrably growing.

The bottom line

Start with the main buying questions in one category. Compare the answers with the store: is the price correct for that variant? Is the delivery time accurate? Once the underlying facts are sound, repeated observations can tell you something about visibility.