Skip to main content
RankRift
Menu
Competitive Intelligence

How To Run A Competitor AI-Visibility Teardown In An Afternoon

A repeatable protocol for finding out who owns the default recommendation in your category.

competitorsteardownresearch

There is a planning meeting on Thursday and someone is going to ask how the category looks in AI answers. You have three anecdotes, two screenshots a colleague sent from their phone, and no way to say whether any of it is representative. This protocol replaces that with a small, honest artifact you can put on a slide and defend when challenged.

Disclosure first: RankRift publishes vendor-neutral criteria for measuring AI visibility, including this spreadsheet protocol, and does not sell a tracker. Apply the commercial-conflict rules to any recommendation from a publisher that does.

What a teardown can and can't tell you in one afternoon

Say this out loud before you start, and put it on the artifact at the end.

A one-afternoon teardown is directional, not conclusive. With a small prompt set and a small number of runs, you are getting a low-resolution read on which brands appear at all and roughly what the source pattern looks like. You are not getting a stable share-of-voice figure, and you are certainly not getting a trend.

What it can support:

  • "These four brands appear in the majority of our category's top-of-funnel prompts and we appear in none of them."
  • "Every answer we sampled cited the same three publications."
  • "Our product appears, but it is framed as the budget option rather than the enterprise one."

What it cannot support:

  • "We hold 12% share of voice." A number that precise from an afternoon's sampling is false precision.
  • "Our visibility rose this quarter." You have one measurement. There is no quarter.
  • "Assistants think Competitor A is a better product." Model output reflects what is written about a category on the web and what a model learned in training. It is a statement about representation, not about quality. Keep that distinction in the artifact, because the room will collapse it if you don't.

It also cannot tell you why, though it will generate good hypotheses. The diagnostic sequence for that lives in why ChatGPT recommends your competitor.

Step 1: freeze a small, fair prompt set

Budget 45 minutes. Aim for 10 to 15 prompts. Fewer than 10 and one odd answer dominates the picture; more than 15 and you will not finish the runs this afternoon.

Build them across buying stages so you are not only measuring one moment:

StageWhat the prompt doesExample shape
Category discoveryAsks for options with no brand named"What tools do [role] use to [job]?"
ShortlistingAsks for a comparison or a best-fit"Best [category] for [constraint]"
Constraint-ledAdds a real buying constraint"[Category] that supports [requirement]"
AlternativesNames a known incumbent"Alternatives to [well-known category leader]"
VerificationAsks about a specific brand's fit"Is [brand] good for [use case]?"

Three rules that make the set fair:

  1. No brand baiting. Do not write prompts that only your product could satisfy. "Best [category] with [the one feature you uniquely ship]" will return you and tell you nothing. A prompt set engineered to flatter the person who commissioned it is worse than no measurement, because it will be believed.
  2. Buyer phrasing, not marketing phrasing. Pull wording from sales call notes, support tickets, and the questions people actually type. Your internal category name is often not the one buyers use.
  3. Freeze the file before you run anything. Save it, date it, version it. If you edit a prompt mid-run, every comparison that spans the edit is broken. Note the change and start a new version instead.

The prompt-set design guidance covers the full construction method if you want to build a set you will keep using.

Step 2: run it consistently and record everything

This is the step where teardowns go wrong, and it goes wrong in a boring way: someone runs eight prompts against one model, gets bored, runs the remaining seven against a different one, and the comparison quietly dies.

The rule that makes the afternoon worth anything:

Every competitor is evaluated from the same prompts, run the same number of times, against the same models, on the same day. No exceptions, including for your own brand.

You do not run separate prompts "for" each competitor. You run one prompt set and record which brands appear in each answer. Competitors are found in the output, not queried individually. Querying a competitor by name and then querying yourself by name produces two different measurements that cannot be compared, and it is the single most common way a teardown misleads the person who ran it.

Practical setup:

  • Runs per prompt: three is the realistic afternoon minimum. Assistant output varies between identical calls, so a single run tells you almost nothing about whether a mention is stable. Three runs at least distinguishes "appeared every time" from "appeared once."

  • Interface: use official model APIs where you have access. Automating consumer accounts or scraping assistant web interfaces breaks provider terms and produces results you cannot publish.

  • Session hygiene: run each prompt in a fresh session with no prior conversation, no memory or personalization features enabled, and consistent region settings. Note what you could not control.

Record one row per run with these columns, every time:

ColumnWhy it exists
Prompt IDTies the run to the frozen set version
Model and version stringBehavior changes between versions; an undated result is unusable later
Run date and timeSame reason
Run number (1–3)Lets you see within-prompt variation
Region / localeAnswers can differ by market
Brands mentionedThe raw observation
Sources citedThe most useful column, see step 4
Full response textPaste it; you will want it during coding

If you record nothing else, record the model version string and the date. A result without those two fields cannot be re-run, compared, or defended, and it becomes worthless the moment a provider ships an update.

Step 3: code the results (mention, framing, sources)

Coding is turning free text into countable observations. Do it with a fixed rubric written before you look at the data, and have one person code everything so the judgement calls are at least consistent.

Code each response on three axes:

  • Mention. Was the brand named at all? Binary. Count per prompt, not per response, so a brand named four times in one answer counts once.
  • Framing. Was it presented as a leading option, an alternative, a niche or budget fit, or a caution? Use four fixed labels and write down the definition of each before coding. This is where a teardown earns its keep, because "we're mentioned" and "we're mentioned as the cheap option" are very different findings.
  • Sources. Which domains were cited or linked in the response, if any.

Then produce two simple counts and stop:

  • Mention rate per brand: prompts where the brand appeared, over total prompts.
  • Framing distribution per brand: how many of those mentions fell into each label.

Report both as raw counts alongside any percentage, so nobody detaches "60%" from "3 of 5 prompts." At this sample size, a percentage without its denominator is a misleading number wearing a lab coat.

Step 4: read the citation-source pattern

The source column is usually the most actionable thing you produce. Mention rates tell you where you stand. Sources tell you what to do about it.

Tally the domains cited across all responses and sort by frequency. You are looking for a handful of patterns:

  • Gatekeeper publications. A small number of sites that appear across most prompts in the category. These are the addresses where being absent is expensive.
  • Community sources. Forums, Q&A sites, and communities appearing repeatedly. Note them as intelligence about where the category's conversation lives. Note also that participating in them is a slow, disclosed, genuinely-useful-contribution exercise, not a posting campaign, and that fake or incentivized participation is detectable and carries real consequences.
  • Owned pages. When a competitor's own documentation or comparison page is cited directly, that is a retrievability signal worth studying.
  • Absent categories. Sometimes the interesting finding is what is never cited: no review platform, no analyst coverage, no independent testing.

The method for turning this tally into a target list is in which sources decide your category.

Step 5: write the one-page summary

One page. Anything longer will not be read by the person who asked for it.

Structure it in this order:

  1. The headline finding, one sentence, hedged accurately. "Across 12 prompts run 3 times each on [model, version] on [date], four brands appeared in more than half of answers; we appeared in two of twelve prompts."
  2. The method box. Prompt set version, number of prompts, runs per prompt, models and version strings, run dates, region, and interface used. This box is not optional. It is what separates the artifact from an anecdote.
  3. The mention table. Brands as rows, raw counts and percentages side by side.
  4. The framing note. Two or three sentences on how you are characterized when you do appear.
  5. The source list. Top cited domains with counts.
  6. The limits paragraph. Directional, not conclusive. Small sample. Single day. Model versions may change. No causal claims. Model output describes how the category is represented in text, not which product is better.
  7. Three hypotheses and the next test for each. This is the part that turns research into a plan.

Attach the raw run log as an appendix or a linked sheet. If someone challenges the finding, the log is your answer.

What to do with it next

Three honest paths, in order of usual value:

  • Re-run it on a schedule before you act on it. One measurement is a point. Two measurements four weeks apart, same frozen prompt set, same run counts, is the beginning of a signal. Most teams skip this and act on the point.
  • Turn the source tally into a workstream. Being cited in the places your category's answers already draw from is slower and more durable than anything else on the list.
  • Decide whether to automate. If you will run this monthly across several brands and models, the spreadsheet stops scaling somewhere around the third re-run. If you will run it twice a year, it never does.

One more thing before you publish anything externally. Internal teardowns are cheap to run because nobody has to litigate them. The moment a competitor comparison leaves the building, in a blog post, a sales deck, or a conference slide, it needs the full fair-benchmarking treatment: published prompt set, disclosed sample size and run counts, model versions and dates, stated limits, and a correction route for any named brand that disputes the result. The rules are in fair benchmarking rules, and they are not optional. A public claim about a named competitor built on an afternoon of sampling is a reputational and legal exposure you do not need. If you are considering publishing, treat that as a decision to run past qualified counsel rather than a formatting question.

More on using assistant output as market research sits on the competitive intelligence hub, and the worksheets and logs live in resources and downloads.

In Plain English

Freeze a small prompt set, run it identically for every brand including your own, record the model version and date on every single run, code mentions and framing with a rubric you wrote first, tally the cited sources, and write one page with a method box and a limits paragraph. The whole thing fits in an afternoon.

This helps a growth lead, strategist, or founder who needs a defensible read on the category before a planning meeting and has no budget or time for a tooling decision. Use it when you have zero baseline. Do not use it as a trend, a score, or evidence that any product is better than another. Do not publish it about named competitors without the fair-benchmarking rules and a legal review.

Next step: run it once this week, then put a calendar reminder to run the identical set again in four weeks. The second run is where it starts being worth something.

Download the competitor teardown worksheet and use the run log as-is for your first afternoon.

Primary-source verification

Checked July 28, 2026. The operational and factual guidance above was reviewed against the primary references below. Provider features, prices, policies, interfaces, jurisdictions, and account-specific results can change; recheck the controlling first-party document and your own account or measured data before acting.