Monitoring what AI models say about your brand: a monthly routine

Why should a fashion brand monitor what AI models say?

Shoppers and wholesale buyers increasingly ask assistants for recommendations, comparisons and background on brands. What those assistants say can be wrong, stale or based on a reseller's page, and the brand will not see it unless it looks. OpenAI itself warns that search results and citations can be incomplete, outdated or incorrect and tells users to open cited sources to verify them.

Monitoring does not require specialist software to start. A disciplined manual routine, run once a month by one named person, produces most of the insight, and it can later be supported by tools if the volume justifies them.

Wholesale buyers, journalists, recruiters and investors also ask these questions. A buyer preparing for a first meeting may ask an assistant about a brand's positioning, price tier and stockists, and an inaccurate answer shapes the first impression before any contact. Including these audiences in the prompt set widens the value of the routine beyond consumer marketing.

What should the monthly routine cover?

Read also

Who owns the customer when an AI agent shops? Brand control in agentic commerce

How should the prompt set be built?

Use real customer language. Draw questions from search console queries, customer service tickets, sales rep notes and reviews. Include brand questions (what is this brand known for), category questions (which brands make a given product), comparison questions (how does one brand compare with another) and buying questions (where to buy, what size to order). For wholesale, add questions a buyer might ask, such as who the brand's retail partners are or what its minimum order is.

Length matters. In an Ahrefs study of 146 million desktop results, AI Overviews appeared for 9.5 per cent of one word queries and 46.4 per cent of queries with seven or more words. A prompt set made only of short keywords will therefore under-sample the conversational questions where generated answers are most common.

• Keep the wording of prompts fixed so month-to-month changes mean something.
• Add new prompts rather than editing old ones, and note the date added.
• Run each prompt in at least two assistants and, for Google, in both standard results and AI Mode.
• Log the market, language and device used.

What counts as a good or bad result?

There is no published benchmark for a healthy mention rate, so avoid inventing one. Instead classify each answer: brand absent, brand present and accurate, brand present with errors, or brand present with outdated information. The most useful measure is the number of factual errors and their causes, since those are fixable. Mention frequency is useful only as a trend within your own prompt set.

Treat single answers with caution. Responses vary between sessions and between users, so a one-off omission is weaker evidence than a pattern across several runs. Record the date and repeat anything that looks important before escalating.

For wholesale and B2B questions, record different things. Note whether an assistant lists the brand's stockists correctly, whether it states the right product categories and whether it describes the brand's positioning in a way the sales team would recognise. Outdated statements about closed stores, discontinued lines or old ownership are common examples worth looking for, and they are usually traceable to a specific old article or directory entry that can be updated or corrected.

How do you fix what you find?

Start with sources you control. If an answer states an old price, a discontinued product or a wrong return policy, find the page or feed that supports it. Update the page, check that it is crawlable and keep the feed consistent with the site. Where the error comes from a third-party site, contact the owner or the reseller and ask for a correction.

Also check access. OpenAI says that sites which opt out of OAI-SearchBot will not appear in ChatGPT search answers, and that a host or CDN should allow its published searchbot traffic. Google says that to appear in AI Overviews or AI Mode a page must be indexed and eligible to show a snippet, and that no special optimisation is needed. A brand that is absent should rule out these basics before assuming anything else.

Read also

How to build an AI strategy for a fashion brand

Who should own it and how should results be shared?

Give ownership to one person in marketing or e-commerce, with a named contact in product data, customer service and legal. Share a one page summary monthly: errors found, errors fixed, pages changed and open items. Include wholesale and sales teams, because buyers may also use assistants to research a brand before a meeting. Review the prompt set each quarter and retire prompts that no longer reflect how customers ask.

1. Appoint an owner and set a fixed day each month.
2. Build and version the prompt list.
3. Run, record and screenshot answers and citations.
4. Classify results and trace errors to sources.
5. Fix sources, recheck and report.

Keep the process lightweight enough to survive. A routine that takes half a day each month and produces a one page summary will be maintained, while one that requires a large report will be dropped by the third month. Store answers and screenshots in a shared folder with consistent names, so that a change in an assistant's behaviour can be shown to colleagues with dated evidence rather than recollection.