LLM Visibility DTC Brand: A Prompt Set You Can Run

What does an LLM visibility DTC brand check actually involve?

Checking llm visibility dtc brand by brand comes down to one repeatable question, asked the same way every time: when a real customer prompt goes into an assistant, does your brand show up in the answer, and how? That sounds simple until you try to do it properly. Assistants don’t return the same answer twice for the same input, an account’s memory and location can quietly change what comes back, and a single screenshot of ChatGPT naming your brand once proves almost nothing about what happens on the next ten tries. A usable check needs structure: a fixed set of prompts, a fixed set of assistants and settings, a place to log every run, and a cadence you actually stick to.

This article is that structure, built as a template you can copy into a spreadsheet today. Every number in it — how many prompts, how many runs per sitting, how often to repeat the cycle — is a working default with the reasoning behind it stated plainly, not a measured industry figure, because no one has published one you could safely treat as a benchmark. Fill in your own brand, your own competitors and your own product categories, and the template does the rest.

Step 1: Build your prompt set

Twenty prompts, split across four categories, is a workable starting point for most single-brand DTC operations. Fewer than that and one odd run — an assistant having a bad day on a single question — swings your whole read of the month; a great deal more than that becomes too slow to run by hand on a monthly cadence, and hand-running it matters, because letting the exact wording drift between cycles breaks the comparison you’re trying to build.

The four categories, with worked examples for a hypothetical skincare brand called “Aster”:

Branded — prompts that name the brand directly, testing whether an assistant recognises and correctly describes it: “What is Aster skincare known for,” “Is Aster a good brand for sensitive skin,” “Where can I buy Aster products.”

Category — prompts a shopper would type without knowing your brand exists, testing whether you surface at all: “Best vitamin C serum for sensitive skin,” “Highly rated skincare brands for rosacea-prone skin.”

Comparison — prompts that pit you against named competitors, testing relative positioning: “Aster versus [Competitor] for sensitive skin,” “Alternatives to [Competitor] that are gentler on skin.”

Problem-solution — prompts framed around the customer’s problem rather than a product category, testing whether the assistant connects your brand to the outcome it solves: “How do I calm redness from a new retinol routine,” “What ingredients should I avoid if my skin reacts easily.”

Write every prompt out in full, exactly as you’ll paste it, and keep that wording locked once you start running the set. A prompt reworded slightly between cycles — “best serum for sensitive skin” versus “top serum for sensitive skin” — can return a genuinely different answer, and you won’t be able to tell whether a change in result came from the market or from your own rewording.

Step 2: Fix the assistants and settings you will test

Pick the assistants your customers are actually likely to use for shopping research before adding every assistant that exists. A general chat assistant, a search-integrated assistant and any AI-generated overview built into the search engine you already rank in cover most DTC shopping research today; add others once the core set is running reliably rather than starting broad and never finishing a single clean cycle.

For each assistant, decide and record three settings before your first run, because all three can change the answer:

• Account state — logged in or logged out. A logged-in session with purchase or browsing history can return a different, more personalised answer than a fresh, logged-out session; note which one you tested.
• Memory — on or off, where the interface exposes the setting. An assistant with memory of past conversations about your category can surface a brand it “remembers” you asking about before, which is a different signal from cold discoverability.
• Model version — where the interface lets you choose or shows one, record it. A model update between cycles can move results on its own, entirely apart from anything that changed about your brand’s presence online, and if you don’t record the version you can’t tell the two apart later.

Running the same prompt set logged in and logged out is more thorough than picking one, but only if you keep the two results in separate columns rather than blending them into a single “did it mention us” tally — mixing account states into one number hides exactly the variation you were trying to observe.

Step 3: Run the set on a cadence and log every result

Run each prompt more than once in the same sitting — three runs is a reasonable minimum — before you draw any conclusion from it. This is the part of the process that most directly answers to the non-determinism of the system you’re sampling: the same prompt, asked twice in a row, can come back with your brand named once and absent the next time, with no change on your end at all. A single run tells you what happened once. Three or more in the same sitting start to tell you whether that result is typical or a fluke.

Do a full cycle monthly as the default, with one extra off-cycle run after anything that could plausibly move the answer — a rebrand, a significant site restructure, a major new product launch, or a large press or influencer moment. Weekly runs are usually not worth the effort for most DTC brands: the underlying pages an assistant draws from don’t typically change fast enough for a week-over-week comparison to mean much, and the extra runs mostly add noise and time cost rather than signal.

Log every run before you interpret any of it, using a sheet with one row per prompt, per assistant, per run. The columns that matter:

Copy the exact sentence into the snippet column rather than summarising it in your own words. A paraphrase loses the specific claim the assistant made, and if that claim is wrong about your product, you need the original wording to know what to correct and where it might have come from.

Step 4: Score what you recorded without overclaiming

Once a cycle’s rows are logged, calculate a simple presence rate: mentions divided by total runs, for each prompt category separately, because branded, category, comparison and problem-solution prompts behave very differently and averaging them together hides that. A hypothetical illustration: if Aster’s five branded prompts each ran three times in a cycle, that’s fifteen runs, and if the brand was mentioned in twelve of them, the branded presence rate for that cycle is twelve divided by fifteen. Do the same arithmetic for each category, and keep the four numbers separate rather than collapsing them into one score.

Compare that rate against your own prior cycles, not against an outside benchmark, because no reliable published figure exists for what a “normal” presence rate looks like across DTC brands, and treating a plausible-sounding number as a target would be inventing a standard that doesn’t exist. Your own trend line, month over month, is the only comparison this method actually supports.

Treat a single run’s absence as noise, not a finding. If your brand appeared in two of three same-session runs for a prompt, the honest read is “usually present, with some run-to-run variation,” not “confirmed visible” or “confirmed absent.” Overclaiming certainty from a small, non-deterministic sample is the single most common way this kind of tracking goes wrong, and it’s an easy mistake to make when a stakeholder wants a clean yes-or-no answer that the method genuinely cannot give.

Keep a running log of month-over-month change per category rather than a single blended trend line, because a brand can gain ground in problem-solution prompts while losing ground in comparison prompts at the same time, and one combined number hides that entirely. A quarterly review that lays the four category trends side by side, rather than reporting one score, tends to surface the actual story: a brand growing steadily in category prompts while stuck at zero on comparison prompts against one specific competitor is a different, more actionable finding than “visibility is roughly flat.”

Step 5: Decide what is worth acting on

Where the brand never appears across a full cycle — zero mentions across every run of every relevant prompt — the next step is diagnostic, not another round of the same test. Check first whether your own site is crawlable and whether your product and brand pages state clearly, in plain text rather than only in images or scripts, what the brand is and does. Then check whether independent third-party pages — reviews, roundups, comparison articles, press coverage — mention the brand at all, because most AI answer mentions come from pages a brand doesn’t control rather than from the brand’s own site, and a total absence from that layer of the web is often the real explanation for a zero score.

Where the brand appears inconsistently — present in some runs, absent in others, for the same prompt — that pattern is usually more useful than either a solid zero or a solid full score, because it points at borderline content: pages that are being crawled and considered but aren’t decisively winning inclusion. That’s a more tractable problem to work on than starting from nothing.

Where a competitor consistently appears and your brand consistently doesn’t, on the same comparison prompts, log the competitor’s cited source alongside your own missing one. Frequently the difference is a specific page type — a comparison article, a detailed spec sheet, an active review programme — that the competitor has and you don’t, which turns a vague visibility gap into a specific, actionable content gap.

How do you adapt this template for a multi-brand or multi-market portfolio?

Run a separate prompt set per brand rather than merging brands into one shared set, because a shared set makes it impossible to tell which brand’s presence moved when the overall number changes. For a portfolio of three brands, that’s three prompt sets, three recording sheets, and three separate presence-rate trend lines, even if the same person runs all three on the same afternoon each month.

For multiple markets, adapt the prompt wording to how a customer in that market would actually phrase the question, rather than translating your English prompt set literally. “Best sunscreen for sensitive skin” and its French or German equivalent aren’t reliably the same search intent once translated word for word, and an assistant’s answer in each market draws on a different pool of local pages. Keep a market’s prompt set stable once it’s set, the same way you would for a single-brand cycle, so month-over-month comparison still holds.

Comparison prompts need the same market-specific care as category prompts. A named competitor with a strong presence in one country may barely register in another, so a straight translation of a comparison prompt built around a domestic competitor can end up testing the wrong rivalry entirely. Rebuild the comparison list per market using whichever competitors actually appear in that market’s category and problem-solution answers first, then lock that market-specific list the same way you locked the original English set.

What breaks if you skip a step or change the prompt set mid-quarter?

Skipping the multiple-runs-per-sitting step is the most damaging shortcut, because a single run gives you a number that looks precise and isn’t. A brand that happened to be mentioned on one run out of what would have been three looks “visible” in a spreadsheet that only ever recorded one attempt, and a decision made on that basis is a decision made on a coin flip that nobody flagged as a coin flip.

Changing prompt wording partway through a quarter breaks the comparison the whole template exists to produce. If “best vitamin C serum for sensitive skin” becomes “top-rated vitamin C serum for sensitive skin” in month three, any change in the result from month two to month three is now unexplainable — it could be the market, or it could be your own rewording, and there’s no way to separate the two after the fact. Lock the wording before the first cycle, and if a prompt genuinely needs to change, retire it and start a new one rather than editing it in place.

Skipping the settings log (account state, memory, model version) has a quieter but similar effect: a model update between two cycles can move your numbers with nothing having changed about your actual visibility, and without a record of which model version each cycle used, you’ll spend time investigating a “drop” that was really just the underlying assistant changing under you.

What are the honest limits of sampling a non-deterministic system?

This template produces a directional read, not a precise measurement, and that distinction matters more than any single number it generates. The systems being sampled don’t return consistent answers, aren’t fully documented in how they select or cite sources, and can change behaviour between one week and the next for reasons entirely outside your control — a model update, a change to how a search engine blends AI answers with traditional results, a shift in which pages an assistant’s retrieval system currently trusts.

A monthly cycle with twenty prompts and three runs each generates sixty data points per assistant, per month. That’s enough to see a trend across several months and to notice a category where you’re consistently strong or weak; it is not enough to treat any single month’s number as a precise measurement of “true” visibility, and it’s nowhere near enough to compare against another brand’s self-reported number, since neither of you is sampling the same way. Report results as a trend with a stated method, not as a score, and the limits of the method stay honest rather than becoming an invented precision nobody can defend.

Who this template is not for

This method assumes a brand with an existing content or marketing function that can act on what the sheet shows — a brand doing $3M to $30M in revenue on Shopify Plus or a comparable paid subscription platform, roughly the floor where a dedicated hour a month of structured tracking is worth someone’s time. A pre-revenue store with no content team yet has more urgent groundwork to do first: getting basic product and brand pages live and crawlable matters more, at that stage, than measuring whether an assistant already knows they exist.

This template only produces measurement, the first half of an AI search visibility programme. What to do about a weak or inconsistent result is a separate, larger body of work that /services/ai-search-visibility is built to cover once you have a baseline worth acting on.