Key points
- Your site is a claim; external sources are the verification of that claim. A model repeats a fact far more readily when it finds it confirmed in several independent places.
- In reviews, the text does the work, not the average score. Specific sentences about use, flaws and durability give a model something to recommend from; a star rating alone gives it nothing.
- Ranked "best X" lists are a convenient source of ready-made answers for models. It is worth knowing which lists actually exist in your category and why you are not on them.
- Making your basic company facts consistent across the web is the cheapest item in this chapter, and the only one you can finish on your own without anyone else’s consent.
After four chapters your own house is in order: you know where models get their knowledge about your store, you have a baseline from the audit, your product data lives in a feed and in schema, and your content is written so an answer can be lifted straight out of it. All of that shares one property, though: it is you talking about yourself. When a model builds a recommendation, it also reaches beyond your domain, into reviews, roundups, trade articles, directories and your product listings on marketplaces. This chapter is about the layer you do not edit, and which often tips the balance: whether the rest of the internet confirms what you say on your own site.
Why other people’s words weigh more than yours
A model has no way to check whether "market leader in its category" is true. What it has instead is a cheap substitute for verification: it looks at whether the same fact repeats across sources that are not yours. A fact confirmed in several independent places is safe to generate. A fact that exists only on your own site stays a claim made by an interested party.
According to industry sources, analyses of citations in assistant answers show the same pattern: on buying questions, models lean far more on third-party sites, meaning roundups, review platforms, comparison services and encyclopedias, than on brand-owned pages. This is not algorithmic prejudice against manufacturers. It follows from how verification works at all.
The practical conclusion is uncomfortable. A mismatch between your site and the rest of the web is not neutral. If a trade directory from five years ago still describes a different offer, and your business listing still carries a pre-rebrand name, the model gets two versions of the truth and picks one. You do not get to choose which.
Reviews: the model reads text, not stars
An average score is a single number and very little follows from it. What matters to a model is the body of the review, because that is where product attributes come from: who it suits, what tends to break, how it holds up in daily use, whether sizing matches the chart. A hundred concrete sentences saying a cream does not irritate sensitive skin count for more in a recommendation than moving an average from 4.6 to 4.8.
In practice that means three things. Collect continuously, after every purchase, rather than in a once-a-year push, because recency is visible. Ask for specifics instead of general satisfaction, because "great, recommended" gives nobody a product description. Reply to negative reviews, because your reply is also text, and it also describes how your customer service actually behaves.
Do not treat unearned ratings as a shortcut, because the risk sits in several layers at once. In the EU, the Omnibus Directive, in force in Poland since 1 January 2023, requires traders who publish reviews to state whether and how they verify that those reviews come from people who actually bought the product, and faking reviews is an unfair commercial practice pursued by the Polish competition authority (as of July 2026). On top of that comes the layer people forget: accounts of such practices stay online, on forums and in the press, which is exactly where a model assembles its picture of a brand.
Rankings and roundups: a ready-made list is the easy path
When someone asks for "the best X for Y", a model can either compare a dozen offers itself or reach for a list somebody has already built. According to industry sources, ready-made roundups and rankings on third-party sites are among the most frequently used sources in buying answers. If five such lists exist in your category, they largely decide who shows up in the recommendation.
The work here is slow and entirely manual. Write down the lists that genuinely exist: trade publication rankings, comparisons on category portals, creator roundups, association picks, regional media selections. Check which ones omit you and find out why. In most cases the reason is mundane: the author did not know your product existed, or had no data to describe it. A plain outreach message with the full set of facts fixes that, meaning specs, images, price, availability and return terms, and in some categories a review unit.
What not to do: do not buy placements in "rankings" that exist only in order to sell placements. You will recognise them by orderings that shift for no visible reason and evaluation criteria described nowhere. Such a list is a weak source for a model too, so you are paying for a mention in a place that convinces neither a human nor a machine.
Mentions and citations in trade media
A mention is not the same thing as a link, and it does not need to become one. What counts is your brand name appearing repeatedly in a credible context: an expert comment in a trade piece, your own numbers quoted in an article about the market, an implementation write-up, a podcast appearance, coverage of a conference talk. A link is a bonus, not the price of entry.
The selection criterion is not reach but credibility inside your category. Ten mentions on sites that publish anything arriving by email weigh less than one in a place the industry actually reads. For a store, the easiest way in runs through data only you hold: seasonality in your category, the most common reasons for returns, how basket composition shifted across the year. Editors look for exactly that kind of material, and it is the part that gets quoted onward.
Consistency of your basic company facts
This is the most underrated item in the chapter, because it needs nobody’s permission, depends on neither editors nor customers, and can be finished in a single afternoon. The point is simple: your basic company facts should read identically everywhere they appear.
| Where to check | What must match | Typical drift |
|---|---|---|
| Maps and search business listing | Name, address, phone, hours, business category | Pre-rebrand name, address from before the move |
| Trade directories, chambers, supplier lists | Description of activity, categories, contact details | An offer description from years ago, products you no longer sell |
| Social media profiles | Brand name, one-sentence description, website address | Three different descriptions of the offer across three networks |
| Marketplace seller accounts | Seller name, brand name, manufacturer details | Brand spelled differently than on your site, with a hyphen or abbreviation |
| Site footer, terms, registry data | Legal entity name, tax ID, registered address | A different legal name in the terms than in the footer |
| Press materials and third-party articles | One sentence about what you actually do | Every piece describes the company slightly differently |
The rule: one set of facts, identical everywhere. Every inconsistency forces the model to pick a version, and you do not control that pick.
Marketplaces and comparison sites as a source of product facts
Your product listings on marketplaces and comparison engines are often more visible to machines than the same products on your own store, because those platforms carry high authority and are crawled intensively. For a model they are a convenient source of product facts, and frequently the first one it reaches. If the title, attributes, EAN or specification differ from what you publish yourself, you are building a competing version of the truth about your own product.
Hence a simple rule: same identifiers, same name, same specification across every channel. Images and description length may differ; facts may not. The side effect of drift is expensive, because a model can describe your product correctly and still point the purchase at a marketplace, where it found the complete data set. The same logic governs shopping agents, which read data instead of viewing pages, something we unpack in detail in our piece on how to prepare your store for agentic commerce.
What cannot be bought or rushed
Honestly: this is the one layer in the guide you cannot close with a two-week project. Reviews arrive at the pace of sales, editors refresh rankings on their own schedule, and a mention requires somebody to decide you have something worth saying. The realistic horizon is quarters rather than sprints, and no button shortens it.
Media budget does not shorten it either. Ads inside AI assistants now exist, OpenAI launched them in ChatGPT in 2026, but they appear as labelled sponsored cards separated from the answer. OpenAI states plainly that ads do not influence the answers, that they run on systems separate from the chat model, and that advertisers cannot shape, rank or alter responses (as of July 2026). Media spend can indirectly raise the odds that somebody writes about you, but it does not buy a place inside the recommendation itself.
That leaves the temptation of shortcuts: bought reviews, manufactured mentions, batches of near-identical articles published under one byline. Beyond the legal exposure described above, there is a purely practical flaw. Models are resistant to sheer volume and sensitive to agreement between sources. A hundred identical, enthusiastic texts in weak places do not create agreement; they create a pattern that is easy to spot and hard to undo later.
Rule of thumb: before you spend a cent building new external signals, align the ones that already exist. Fixing stale directory entries, your business listing and your marketplace seller data costs one afternoon, and it removes the contradictions that make a model describe your company as it was three years ago.
Questions
Does review volume matter more than the average rating?
The text matters more than either. Volume and recency build credibility, but it is the sentences about specific use, flaws and durability that give a model something to build a recommendation from. A catalogue with hundreds of ratings and no written reviews is, for this purpose, close to empty.
Is it worth paying for a spot in a "best X" ranking?
It depends what the ranking is. Paid, clearly labelled presence in a credible editorial roundup can be justified like any other advertising, but it does not buy a recommendation inside a model. Lists built purely to sell placements are weak sources for AI systems too, so you end up paying for a mention that changes very little.
Is an unlinked brand mention worth anything?
Yes, though differently from a classic link. Your brand name appearing repeatedly in a credible context helps systems recognise the company as a consistent entity and associate it with the category it operates in. A link adds value by making the source easy to reach, but the mention itself already enters the material a model builds its answer from.
