By KYN AI Advisory Team — AI implementation specialists, Singapore
A marketing agent invents competitor claims because it predicts plausible words, not facts. Asked for "faster than X" or "priced below the market", it writes what sounds right unless it was given a checkable fact. Three controls catch this before launch: a research pass separate from the writing pass, a reviewer that names exactly what is unsupported, and a rule that every competitor claim traces to a source you can open. After launch, track what AI engines say about you.
Where do invented competitor claims come from?
The simplest agent is one model call: topic in, copy out, publish. Teams build it first because it demos well. It is also the design most exposed to this error, because nothing separates what the business knows from what sounds right.
Two failures follow within days:
- Generic prose that could describe any company in the category.
- A false statement about a competitor or the market, such as a price, a missing feature or a ranking, that goes live because nothing was built to catch it.
Both have one cause. The system cannot tell verified fact from prediction.
Why separate research from writing?
Split the job into two steps.
- Research. Pull specific, checkable facts from your own material and from public competitor pages: pricing sheets, documentation, published comparisons. Store each fact with its URL and the date checked.
- Writing. Draft only from that fact set, not from the model's general knowledge of the category.
The test is traceability. Every competitor claim in the final draft must map to a stored fact. If it does not, it is flagged as unsupported and either verified or cut.
What should the agent do when it has no fact?
Say so. Never guess a competitor's price, feature or market position. A quotation agent asks a clarifying question when a currency or quantity is ambiguous and never guesses a number on a client document. Competitor claims need the same rule.
Treat errors as unequal. A confident false comparison that ships costs far more to unwind than a draft held for review. Build the gate to bias toward flagging.
Why is a quality score not enough?
A reviewer that returns "7 out of 10" gives you nothing to fix. It either blocks the draft or passes it.
A reviewer that returns named feedback does the work of an editor:
- "The competitor price on line 4 has no source."
- "The feature comparison is not connected to the campaign message."
The writer fixes only those points. Cap the loop at a small number of rounds. Without a cap, a disputed claim bounces between reviewer and writer with no end. After the cap, the draft goes to a named person, who decides whether to cut the claim or source it.
What is a pre-ship checklist for competitor claims?
- Every competitor price carries a link and the date it was checked.
- Every "only" and "first" claim has been checked against a search, or removed.
- Comparisons name what was compared and on what date.
- Prices on pages you do not control are re-checked before each campaign, not once.
- A person approves anything that names a competitor, before it publishes.
On the legal side, comparative advertising rules differ by country. Check the rule for the market where the ad runs. Where the claim is a price, the source link is your evidence.
How do you check what AI engines say about you after launch?
Pre-ship controls decide what you claim. Visibility tracking shows what AI answer engines tell buyers, independent of your claims.
Run a fixed set of realistic buyer questions against several engines on a schedule. Use questions a stranger would type, not your brand name. For each answer record three things:
- Whether you were mentioned.
- Which competitors appeared, and in what position.
- How the mention reads: a line in a list of ten, or the direct answer.
Keep results per engine. Engines disagree, and an average hides a gap where one engine never names you. Compare the query set with your search console data. A query that earns impressions but matches no tracked prompt shows a blind spot in your prompt list.
KYN runs an AI search check at about S$0.05 per check, and our own set is 22 checks.
Which graph fits the pipeline?
A planner sets the brief. A researcher extracts the fact set. A writer drafts from it. A reviewer either passes the draft or sends it back with named feedback, up to a cap. A plain sequential chain cannot express that loop. Narrow nodes let each step do one job and let the reviewer act as a real check.
This is also why publishing without a person reading every line can be workable: an adversarial review does the editor's job, and a monitoring step watches the result. The exception is anything that names a competitor or states a price. A person approves those.
What kinds of competitor error do agents make?
The errors fall into six groups. Naming them lets a reviewer check for each one.
| Error type | Example (illustrative) | How to catch it | |---|---|---| | Invented price | "Competitor X starts at a lower monthly fee" with no source | Require a price link and check date | | Stale price | A price that was true last year | Re-check before each campaign | | Missing feature | "X has no reporting module" when it does | Check the vendor's own documentation | | False ranking | "The most used tool in the region" | Remove, or cite a named survey | | Merged products | Features of two plans or two products blended | Name the plan in the fact record | | Borrowed claim | A competitor's marketing line restated as fact | Quote it as their claim, with the page |
The model is not lying. It has no record of where a claim came from, so a claim from a rival's page, an old blog post and its own guess look identical to it.
What does a fact record look like?
Store each competitor fact as a row, not as text in a prompt. A row has six fields:
- Competitor and product or plan. For example, the name and the plan tier.
- Claim. One statement, such as "includes an API on the entry plan".
- Source URL. The page you opened.
- Date checked. The day a person or the research step opened it.
- Checked by. The agent step or the named person.
- Expiry. A review date, set shorter for prices than for features.
The writing step may use only rows that have not expired. An expired row is treated as missing: the draft either leaves the claim out or holds for a refresh.
A worked example
This example is illustrative. The products and numbers are invented.
A campaign brief asks for a page comparing your reporting tool with a rival's. The research step pulls four facts from the rival's public pages and your own documentation. The writer drafts six comparison sentences.
The reviewer returns this feedback:
- Sentence 2 says the rival "charges extra for exports". No stored fact supports it. Source it or cut it.
- Sentence 4 says the rival "has no mobile app". The stored fact is dated 14 months ago, and its expiry has passed. Re-check.
- Sentence 5 says your tool is "the fastest to set up". No stored fact supports it, and the claim is a ranking. Cut it.
The writer fixes the three points. Sentence 2 turns out to be true, and a person adds the source link. Sentence 4 is now false, because the rival launched an app, and it is removed. Sentence 5 is cut. Three of six sentences needed action. A one-pass agent would have published all six.
Who approves what?
| Content | Agent alone | Needs a named person | |---|---|---| | Draft in a review queue | Yes | | | Statements about your own product that match your documentation | Yes, with a check | | | Any statement that names a competitor | | Yes | | Any price, discount or ranking | | Yes | | Publishing to your site or ad account | | Yes | | Re-checking a source and updating a fact record | Yes | Review weekly |
The rule is short. The agent prepares. A person approves anything that could be wrong about someone else.
How do you test the pipeline before it goes live?
Build a test set of 20 briefs that invite a competitor claim. Include some where no fact exists and some where a fact exists but is out of date. Then run the pipeline and count:
- Claims published without a source (the target is zero).
- Claims flagged that were in fact sourced (false alarms).
- Briefs where the agent said it had no fact and held the draft.
- Time a person spent clearing the queue.
Run the test again whenever you change the prompt, the model or the source list. A new model can change the rate of unsupported claims.
What about legal and platform rules?
Comparative advertising rules differ by country, and ad platforms have their own policies on naming competitors. Ask your lawyer about the markets you advertise in. Keep the evidence file: the source link, the date checked and who approved the claim. If a competitor disputes a statement, the file is your answer.
What edge cases catch teams out?
- A competitor changes its name or plans. Fact rows tied to an old plan name become unmatched. Track plan names as separate values.
- Prices behind a login or a quote. If the competitor does not publish a price, do not estimate one. Write "not published" and name the date you checked.
- Paywalled or blocked pages. The research step may fail to read a page. It must report the failure, not fill the gap from memory.
- Review sites. A rating on a third-party site is a claim by that site. Quote it with the site's name and date.
- Translated pages. A fact read in another language needs a person who reads that language to confirm the wording.
- Old pages that stay indexed. Your own earlier articles may hold a claim that is now false. When a fact row is updated, search your site for the old wording and queue those pages for a person to review.
- Screenshots as sources. Save the screenshot with the URL and date, because the page may change before anyone checks the claim again.
Where KYN fits
KYN builds marketing agents on cloud infrastructure in the client's own account. Agents work from their own restricted accounts, every action is logged with the actor, and anything published or sent under your name waits for a named person. Our article engine costs about S$0.06 per article.
Before any build, KYN spends the first 14 days reading how your team writes, checks and publishes content today, and listing the sources your claims rely on. Delivery of the full pipeline comes in under 90 days from there. See marketing systems and the AI search check.
Frequently asked questions
Can an AI agent write competitor comparisons safely?
Yes, if it drafts only from stored, sourced facts and a person approves anything that names a competitor. Without that, treat every comparison as unverified.
How often should competitor prices be re-checked?
Before each campaign that cites them, and on a schedule for evergreen pages. Competitor pages change without notice, so a price checked last quarter is not evidence this quarter.
What is the fastest control to add first?
A rule that no competitor claim ships without a source link and check date. It needs no new tooling and removes the main route for invented claims.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
