Skip to main content
RankRift
Menu
Risk & Brand Safety

When A Competitor Comparison In An AI Answer Becomes A Legal Problem

Model output isn't an advertising claim. What that means for your options, in general terms.

legalcompetitorsbrand-safety

Scope note. This article is general educational and measurement guidance, not legal, tax, financial, security, or platform-specific advice. Current law, first-party platform rules, the facts of the real case, and qualified professional advice control any consequential decision. This post is general information, not legal advice. It does not state what the law requires or permits anywhere, it is not a substitute for advice from a qualified lawyer in your jurisdiction, and reading it creates no professional relationship of any kind. If a specific incident matters to your business, take it to counsel with your documentation, not to a blog post.

An assistant told a prospect that a competitor handles something your product handles too, or that your product lacks a certification you hold. Someone forwarded the screenshot with a note asking whether you can send a letter. Before anyone drafts anything, it is worth understanding what kind of problem this usually is, because in most cases it is a source-correction problem wearing the costume of a legal one.

The difference between model output and an advertising claim

The instinct behind "can we send a letter" comes from a familiar setting: a competitor publishes a comparison page that misstates your product, you know who wrote it, and there is a named advertiser standing behind the claim. The frameworks marketers reach for grew up around that setting.

An assistant's answer looks superficially similar and is structurally different in ways that matter to anyone assessing it:

  • There is often no author in the ordinary sense. The sentence was assembled at request time and may never have existed in that form before, and may never exist again.
  • It is frequently not reproducible. Ask twice and you may get two different characterisations, which complicates any attempt to establish what was said.
  • The audience may be one person. Unlike a published page, an answer generated in a private session has no fixed readership you can point to.
  • The claim usually traces to something else. Very often the model is restating, compressing, or garbling material that exists somewhere retrievable.
  • Multiple parties are involved. The model provider, whatever sources were retrieved, and in some cases a competitor's own published material.

How any of that maps onto advertising rules, unfair-competition frameworks, or defamation doctrine in your jurisdiction is a legal question with jurisdiction-specific answers, and it is genuinely unsettled in places. We are not going to guess at it, and you should be wary of anyone in this category who does. What we can say is operational: the shape of the problem is different enough that the reflexes built for competitor advertising disputes are a poor first tool.

Why most cases are source problems

Work backwards from the output and you will usually find something concrete. In practice the false comparison tends to trace to one of these:

What you seeWhat is often underneathWhere the fix lives
"X supports this, you do not"An outdated third-party comparison page or directory entryThe publisher of that page
A wrong certification, capability, or availabilityYour own documentation, which never states it plainlyYour site
A wrong category or an odd peer setWeak entity clarity, so the model places you by resemblanceYour site and reference sources
A stale price model or tierCached third-party coverageThe publisher, plus a clearer canonical page
A confident claim you cannot source at allCompression across several sources, none of which said itSources plus repeated measurement

The practical consequence: the fastest route to changing what an assistant says is almost always fixing or getting corrected the material it is drawing on, not confronting the party that generated the sentence. That is the sequence in the correction playbook, and it is worth exhausting before anything escalates.

It also has a cost profile that legal action does not. Contacting a review site with a dated correction is cheap, fast, and repeatable. It can be measured. Nobody has to approve a budget for it.

One thing it is not: a guarantee. No sequence of corrections controls what a model outputs. Sources change what is available to be retrieved, and that is the extent of the influence anyone has.

The situations that are genuinely different

Some incidents do sit in a different category, and recognising them early matters because the documentation burden is higher and the window can be shorter. Without characterising any of these as legal claims, these are the patterns where teams commonly involve counsel:

  • Reproducible, specific, and damaging. The statement recurs across runs, is a checkable factual assertion rather than a matter of opinion, and touches something material such as safety, licensing, solvency, or regulatory standing.

  • Traceable to a competitor's own published material. The model is faithfully repeating a claim a named competitor put in public. The dispute is with the competitor's publication, and that is a familiar setting again.

  • Evidence of deliberate manipulation. Coordinated fake reviews, astroturfed community posts, or content published specifically to influence what assistants say about you. Documentation matters more than confrontation here, and the platforms hosting that material generally have their own detection and enforcement interest.

  • Trademark or brand-identity confusion. Your name attached to another company's product or claims.

  • Regulated-industry exposure. Categories where a false statement about your product creates a compliance problem independent of any dispute.

  • Scale. The same false statement appearing across multiple models and persisting through several correction attempts.

Even in these cases the first action is not a letter. It is documentation.

Documenting properly, in case it escalates

Screenshots are weak evidence and they are what most teams have. A screenshot shows one output at one moment, with no model version, no prompt provenance, and no indication of how often the statement occurs. If an incident does escalate, the quality of your record is the difference between an assertion and a demonstration.

Build the record like a measurement run:

  1. Capture the exact prompt. Verbatim, including phrasing quirks. Note where it came from and whether a real buyer would ask it that way.

  2. Record the model, version, and access date for every run, and keep them consistent across the record.

  3. Re-run it, repeatedly and across days. Establish how often the statement appears rather than that it appeared once. A statement occurring in a minority of runs is a materially different fact than one occurring in nearly all of them.

  4. Vary the phrasing deliberately. If it only appears under one unusual wording, that belongs in the record too. Suppressing it damages your own credibility later.

  5. Preserve full outputs, not crops. Including any citations shown, with the cited URLs captured and dated.

  6. Trace and archive the sources. Save the third-party pages the claim appears to come from, with access dates, before anyone corrects them.

  7. Use official APIs. Records built from documented interfaces are versioned, dated, and repeatable. Records built from driving a consumer interface are none of those things.

  8. Keep it contemporaneous and unedited. Log as you go, note what you did not find, and date everything.

  9. Write down the harm you can actually evidence, and be equally clear about what you are inferring.

The same discipline that makes a competitor comparison defensible applies here. Our fair benchmarking rules cover holding prompts, models, and dates constant so that a difference means something, and that standard is exactly what a lawyer will want if this goes further.

When to involve counsel

Counsel is not the last resort after everything else fails. Bring them in early, in an advisory posture, when any of these are true:

  • The statement is reproducible, specific, factual, and material.
  • You operate in a regulated category where a false statement creates independent exposure.
  • The claim traces to a named competitor's own publication.
  • You are considering any external communication about the incident, including a public post.
  • Multiple jurisdictions are involved, which is the normal condition for anything on the internet.
  • You are contemplating contacting a model provider formally rather than through a standard feedback route.

Bring the documentation package to that conversation. What counsel needs is the reproduction record, the source trail, the timeline, and the harm evidence, all dated. What they cannot work with is a screenshot and an urgent tone.

A note on escalating first. Sending a threatening letter as an opening move is expensive, it can harden the position of a publisher who would have quietly corrected a page, and it creates a written record of its own that you may not want. It is also poorly matched to a problem where the likeliest cause is an outdated third-party page and the likeliest fix is an email to its editor. Advisory counsel early, aggressive action rarely and deliberately.

What to do in the meantime

While the record builds and, if warranted, counsel considers it:

  • Fix your own canonical sources first. State the disputed fact plainly, in one place, with a date. This is entirely within your control and it costs nothing to start today.
  • Request corrections on third-party pages with a polite, evidenced, dated note. Most publishers correct factual errors when handed the evidence.
  • Keep measuring. Track the statement's frequency over time so you can tell whether anything you did coincided with a change, without claiming it caused one.
  • Brief sales and support. Give them an accurate, non-defensive response for prospects who arrive with the wrong information. This handles the immediate commercial damage while everything else is slow.
  • Do not respond in kind. Publishing counter-claims about the competitor, or seeding content designed to influence what assistants say, converts a source problem into a conduct problem with its own exposure.
  • Log the incident internally, dated, so that a repeat six months later is recognisable as a pattern rather than a surprise.

Our approach to disputes about anything we publish is set out on corrections and disputes, and the same posture is the one we recommend when you are on the receiving end: evidence, dates, and a route to reply.

In Plain English

What this means: an assistant saying something false and unflattering about you is usually a problem with the sources it drew on, not a dispute with an advertiser, and the fastest fixes are correction requests and clearer documentation of your own.

Who this helps: brand, comms, and marketing leads holding a screenshot and a demand for action, and anyone who needs to explain internally why a letter is not step one.

When to use it: the day the screenshot lands, before any external communication goes out. Broader brand-exposure material sits on our risk and brand safety hub.

What to do next: build the reproduction record with model versions and dates, trace and archive the underlying sources, fix your own canonical pages, and take anything reproducible, specific, and material to a qualified lawyer in your jurisdiction. Nothing in this post substitutes for that conversation.

Download the brand incident playbook.

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.