By KYN AI Advisory Team — AI implementation specialists, Singapore
A sales bot answers from the knowledge base it was last given. If a competitor changed its pricing last week, or a feature you call a differentiator shipped elsewhere two releases ago, the bot does not know. It states the old fact with the same confidence as a current one, and the prospect cannot tell the difference. The fix is not a smarter bot. It is competitor data with a date and a source on every fact, a scheduled refresh, and a rule that old facts are flagged, not stated.
What is competitor-intelligence blindness?
It is the failure of a sales agent to distinguish a fact verified last week from one verified at launch. The bot does not hedge. A reviewer reading the transcript after a lost deal usually cannot tell either.
Typical examples:
- a pricing tier quoted that the competitor retired;
- a comparison table showing a missing feature that now exists;
- a claim about a certification, region or integration that changed;
- a case study cited for a product line that no longer exists.
The damage is not just a lost deal. A wrong claim about a named competitor can embarrass the salesperson in front of the buyer, who checked.
Why does it happen?
Three architectural causes.
- Built once, trusted forever. The knowledge base is loaded at launch and treated as ground truth until a person notices a mistake. Nothing in the system asks "is this still true?"
- No dates on facts. Retrieval returns text, not text plus "checked on". A fact verified 18 months ago looks the same as one verified yesterday.
- No owner. Nobody is assigned to re-check competitor claims. Marketing owns the page. Sales owns the bot. Neither owns the freshness.
Competitor information is also information about a system you do not control. Treat it as you would any external data feed: it changes without notice.
How do you detect stale competitor data?
Check four things on a schedule.
- Age. For each competitor fact, the date it was last checked and by whom. List everything older than a threshold you set, such as 30 days for pricing and 90 days for features.
- Source. Each fact links to a public page. If the page changes or disappears, flag the fact.
- Disagreement between sources. Compare the competitor's pricing page, a marketplace listing and a recent third-party review. If they differ, a person resolves it.
- Questions the bot cannot answer. Log the buyer questions the bot deflected or answered with low confidence. The topics you never thought to test are where blindness lasts longest.
Single pass-or-fail checks hide problems. A fact can look fine against one reference and be wrong against another. Do not average away disagreement.
What design rules keep a bot honest?
Date every fact. The bot may state a competitor fact only if it has a source and a check date inside the allowed age. Beyond that it says "I don't have a current figure" and passes the question to a person.
Prefer a flag to a guess. In our reconciliation agents, a payment that does not match within a tight tolerance is flagged for review, not auto-resolved. A confident wrong answer costs more than an open item. The same applies to competitor claims.
Require a second signal to clear a flag. To mark a stale fact as current again, require an independent confirmation, such as the live pricing page checked today, not just a person's recollection.
Trace every claim. A statement about a competitor in a draft must link to a stored fact, not to what the model remembers from training. If no stored fact exists, the draft does not make the claim.
Cap revision loops. If a draft keeps failing review for unsupported claims, hand it to a person after a fixed number of rounds.
Keep unproven updates out of live conversations. Validate new competitor data in a test copy before the bot sees it.
How should the refresh pipeline work?
- Collect. Fetch each competitor page you track, on a schedule.
- Compare. Detect a change against the stored fact.
- Propose. Write the change as a proposed update with the old value, the new value and the source link.
- Approve. A named person confirms. Money, pricing and public claims wait for a person.
- Publish. The new fact enters the knowledge base with today's date.
- Log. Record who approved, when, and what changed.
If you use public pages, respect their terms. The check should read prices that anyone can see. It should not log in or work around access controls.
Where AI search tools describe your competitors, run the same check there. Compare how several assistants describe you and them on realistic buyer questions, on a schedule. Our guide to how marketing agents hallucinate competitor positioning and the one on keeping AI-generated battlecards updated cover that side.
How do you measure the damage?
You can count it partly. For deals lost, read the last 30 conversations where a competitor was named and mark every claim the bot made about them. Check each against today's public information. The count of wrong claims, and what share of lost deals involved at least one, gives you a baseline. We do not have a published rate for how fast competitor claims go stale in the Singapore market, so measure your own.
What does a dated fact register look like?
The register is a table the bot reads from. It replaces free text in the knowledge base. The rows below are illustrative, for an invented competitor called "Vendor X".
| Fact | Value | Source | Last checked | Max age | Status | |---|---|---|---|---|---| | Entry plan price | S$99 per user per month (illustrative) | Vendor X public pricing page | 3 Oct | 30 days | Current | | Single sign-on included | Enterprise tier only (illustrative) | Vendor X feature page | 12 Aug | 90 days | Stale: bot must not state it | | Data centre in Singapore | Yes (illustrative) | Vendor X trust page | 28 Oct | 90 days | Current | | Free trial length | 14 days (illustrative) | Vendor X sign-up page | 30 Sep | 30 days | Due for re-check |
The "Status" column is computed, not typed. The system compares today's date with the last-checked date and the maximum age. A row past its limit changes state on its own, and the bot stops stating it. Nobody has to remember.
Choose the maximum age by how often the fact moves. Prices and promotions change often, so 30 days is a sensible starting point. Features and certifications move more slowly, so 90 days. Company ownership and office locations can wait longer. Adjust after the first quarter, using the number of facts that changed at each re-check.
How should the bot answer when a fact is stale?
Consider this exchange. A prospect asks, "Is Vendor X cheaper than you for ten users?"
Weak answer: "Vendor X charges S$99 per user, so you would pay S$990 a month with them." The figure may be six months old. The buyer will check it in two minutes.
Better answer: "I do not have a current price for Vendor X that I can quote. Their pricing page is public, and I can have a colleague send you a side-by-side comparison with today's figures." The bot declines the number, offers a next step and creates a task for a named person.
The better answer costs nothing in trust. A buyer who asks a pointed comparison question usually values a precise answer over a fast one. The rule behind it is short: no dated, sourced, in-limit fact, no claim.
Which competitor claims are riskiest?
Rank claims by the damage if they are wrong, and set tighter review on the top of the list.
- Price and discount claims. Easy to check, so a wrong one is exposed fast.
- Claims that a competitor lacks a feature. The competitor's sales team will correct these in front of your buyer.
- Compliance and certification claims. A false statement here can reach legal or regulatory territory. Route every such claim to a person.
- Customer and case-study references. A reference to a named customer of a competitor should never come from the bot.
- Market position claims. "Larger than", "more popular than" and similar need a sourced figure or should be left out.
What edge cases catch teams out?
- Regional pricing. A competitor may price differently in Singapore, Hong Kong, Indonesia and Malaysia, and in different currencies. Store the region and currency with the fact, and quote only the one the prospect is in.
- Promotional prices. A time-limited discount looks like a price change and then reverts. Store the promotion end date, or record list price and promotion separately.
- Pages behind a login. If the figure is only visible after sign-in, you cannot check it from a public page. Mark it "not publicly verifiable" and keep it out of the bot.
- Rebrands and acquisitions. A competitor that changes its name leaves your register pointing at a dead page. Treat a failed fetch as a flag, not as "no change".
- Sales reps overriding the register. A rep who pastes a claim into the conversation bypasses all of this. Log manual edits with the actor, and route them through the same approval step.
What does a weekly review take?
Plan on 30 to 45 minutes for a team tracking five competitors with about 12 facts each (illustrative). The reviewer opens the list of rows past their limit, checks each against the live page, and approves or rejects the proposed update. Then they scan the log of questions the bot deflected and add any recurring topic to the register. Over a quarter this shortens, because the register covers more of what buyers ask.
Where KYN fits
KYN does not sell a product dedicated to competitor intelligence. We build sales and marketing agents in your own cloud account with rules, dated facts and approval gates, and this refresh pipeline is a pattern we design into them. Agents have their own restricted accounts, every action is logged with the actor, and outbound messages and public claims wait for a named person. A live example: an insurance brokerage's outreach carries one checkable first line per email, with the source shown to the reviewer. See sales systems and marketing systems.
The first 14 days are a review; a full build ships in under 90 days. Pricing is a one-off build fee and a monthly retainer of at least twelve months, quoted after the review.
Frequently asked questions
Why do sales bots quote outdated competitor information?
Their knowledge base is built once and not refreshed. Facts carry no check date, so the bot cannot tell fresh from stale, and no one owns the re-checking.
How often should competitor pricing in a bot be refreshed?
Set it by how fast each fact changes. Weekly or monthly for pricing, quarterly for features, as examples. Whatever the interval, every fact should carry a check date and the bot should flag any past its limit.
Should the bot ever state a competitor fact without a source?
No. If it has no stored, dated, sourced fact, it should say it lacks a current figure and hand the question to a person.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
