What AI says your product is bad at: turn objections into a review backlog

To find out what AI says your product is bad at, ask the engines directly why a buyer might not choose you, then read the recurring answers as a review backlog. Do not treat them as a verdict. An engine asked for downsides will usually produce some for any brand, so the useful signal is which objections come up first, how they compare with your competitors, and whether the quote behind each one points to something you can check, fix or explain. This guide shows how to triage those objections, route them to product, marketing or support, and keep the backlog honest.

In short

• A direct objection question is built to produce downsides, so almost every brand has some. Compare objections with each other, with competitors and with earlier studies instead of reading any single one as bad news.
• Separate objections (what engines say when asked for downsides) from unsolicited weaknesses (what they volunteer in answers to your ordinary tracked prompts). They answer different questions.
• Every backlog item needs a quote, a triage call (true, outdated, unclear or category-wide) and one owner.
• Objections describe what an engine said. They do not show why it said it or how many real buyers hold the view.
• Review on a rhythm, and look at what is new or rising before you look at what is biggest.

What is an AI product objection?

An AI product objection is a reason an AI engine gives, when asked directly, for why a buyer might not choose a brand. It is a claim in an answer, not a customer's opinion and not a fact about your product.

In DiscoveredBy, objections come from a separate weekly study, not from your tracked prompts. Each chat engine on your plan, plus ChatGPT (app), is sent one fixed question about your brand and about each active tracked competitor (Google AI Overviews, Google AI Mode and Gemini (app) are not asked): a buyer is considering the brand and wants to understand the downsides, so what are the main objections, listed from most to least significant, with sources where possible. Another AI model reads each answer and lists the objections it states, and every objection keeps the sentence from the answer that states it. See the Objections docs for the exact mechanics.

That design has a consequence you must carry through the whole exercise. The question asks for downsides, so an engine will usually list some, even for a well-regarded brand. The study tells you which objections engines reach for first when asked. It does not measure how negative engines are about you in ordinary answers.

Enlarge image

Objections or unsolicited weaknesses: which one are you looking at?

Objections are downsides an engine gives because you asked for them. Unsolicited weaknesses are downsides an engine gave in an answer to a question that did not ask for them. Mixing the two overstates or understates the problem.

Brand reasons reads the answers to your tracked prompts and counts the reasons they give for or against each brand, using thirteen fixed labels. Tracked prompts do not usually ask for downsides, so the weaknesses there are what engines volunteer. Objections asks directly, every week, and groups the answers by meaning rather than by fixed label. Study answers are kept apart from your tracked-prompt answers, so neither feeds the other.

Use this to decide where a concern belongs:

A concern that shows up in both deserves earlier attention, because it appears in answers to the questions you track as well as in answers to a direct request for downsides. A concern that shows up only in the objection study may simply be an engine listing what it can when prompted. Neither is proof of what buyers think.

How do you read the objections screen?

Read four things for each objection group: prominence, raised by, also raised about, and change. Then open the quotes.

• Prominence is the average, across the engines that answered for your brand, of each engine's score for the group, from 0 to 100. An engine's score is set by where its quote sits in the answer: the earliest kept objection scores 100 and each later rank loses 10 points, and if an engine has several objections in one group, its highest score counts. An engine that answered without raising the group scores 0.
• Raised by is how many of the answering engines raised the group, out of the engines that answered.
• Also raised about (your own brand only) is how many competitors had at least one engine raise the same group about them, out of the competitors with at least one engine answering. A high count points to a category-wide objection rather than one aimed at you.
• Change compares with the previous study that asked the same question version in the same language, using only engines that answered in both. The labels are New, Rising or Falling (a move of 15 points or more), Steady, and No comparison.

Prominence is relative. Compare objections with each other and across studies; do not read a high number as bad news on its own. Objections that no engine raised this time but a compared engine raised last time move to a separate "No longer raised" list.

The grouping and its labels are an AI model's judgement. Two similar objections can land in separate groups, or a group can be broader than you would draw it. The quotes under each group are your check on that.

How do you triage an objection?

Open the quotes behind each group and give every one of them a triage call before anyone is assigned work. The call decides whether the response is a product fix, a content fix, an answer for support, or no action.

Work through these questions in order:

1. Is the claim about a real thing? Compare the quote with your product, pricing and documentation. Is it true, partly true, outdated, or wrong?
2. Is it category-wide? If several competitors have the same objection raised about them, it may be a normal concern for the category. That changes the response from "fix" to "address plainly".
3. Is it new or rising? Those deserve attention sooner than a steady one.
4. Does a source sit behind it? A linked source shows where the engine placed its citation. It does not check that the page says what the objection says. An objection with no linked source shows "No source was tied to this objection."
5. Can you act on it? Some objections are decisions about the product roadmap, and some are simply out of your control.

Source links depend on what each engine records. Claude records the text it cited but no position in its answer, so objections from Claude answers show no sources. Perplexity, Grok, ChatGPT (app) and Gemini (API) record positions in different ways, and the docs describe the gaps, such as Gemini (API) positions drifting for non-Latin scripts. Read the engine notes in the Objections docs before you lean on a source link, and check the engine list on Engines and measurement.

Enlarge image

The triage rubric

Use this table when you review a group. Fill one line per objection group.

Where a claim depends on a fact you want to lock down, the separate fact check study checks what engines say about your brand against facts you approve. That is a different workflow from this one, so keep the two apart.

A copyable review backlog template

Copy this into a spreadsheet or your tracker. One row per objection group per review.

Review date:                  Study date:               Reviewer:
Question version/language:    Engines answering:        Compared engines:

Objection group (label):
Prominence:        Raised by (x of y):        Also raised about (x of y):
Change status:     New / Rising / Falling / Steady / No comparison / No longer raised
Also in Brand reasons? (label, direction, rate, provisional?):   yes / no

Best quote (engine, rank, sentence):
Linked source (URL, or "none"):
Does the source page actually say this?   yes / no / not checked

Triage call:    true-fixable / true-explainable / outdated / wrong / unclear / category-wide / not actionable
Evidence I checked (page, doc, roadmap item):
Owner:          Product / Marketing / Support / Other
Proposed action (a hypothesis, not a promised outcome):
Due / next review date:
What would tell me it worked (next study, same engines):

Worked example: Quillstone's first three reviews

Quillstone sells document-review software to legal and compliance teams. Its objection study has three engines answering for Quillstone this week. The team reviews the top groups.

The arithmetic: integrations is (90 + 100 + 100) ÷ 3 = 96.7. Pricing is (100 + 80 + 0) ÷ 3 = 60.0. Setup is (80 + 0 + 0) ÷ 3 = 26.7. If last week's prominence for integrations was 60.0 and for pricing 66.7 over the same three engines, the changes are +36.7 and -6.7. Only integrations moved by 15 points or more, so only it is Rising.

The team triages each one.

• Integrations. It is rising, raised by every engine, and shared by two of three competitors. The quote in Engine B's answer says Quillstone "only connects to a few document systems". The team checks its integration list and finds the claim is partly outdated: two connectors shipped last quarter but the integrations page still lists the old set. Triage call: outdated, plus true-fixable for the systems that are still missing. Owners: marketing updates the page (see the integration page review checklist), product decides which missing connectors to plan.
• Pricing. Raised by two engines, steady, and only one competitor shares it. The quote says add-on fees are hard to predict. The team compares it with its pricing page and finds the fee schedule is scattered over two pages. Triage call: unclear. Owner: marketing, to consolidate the fee explanation on one page.
• Setup. New, one engine, no competitor shares it. It is the weakest evidence. Triage call: not actionable yet. The team records it and checks next week rather than assigning work off a single answer.

Notice what the team did not do. It did not claim the integration edit will fix the objection, and it did not treat the score as a measure of how buyers feel. It wrote a hypothesis and set a check on the next study.

How often should you review, and what should you compare?

Review after each weekly study, starting with New and Rising items. Studies run automatically once a week, on Wednesdays at 05:00 UTC, for projects whose owner's plan includes Objections, unless one is already running or one was created in the last 6 days, so a study you start yourself early in the week can make the Wednesday run skip that project. Owners and editors can also press Run now, with a 24 hour cooldown, and viewers can read the screen but not start a study.

Only compare like with like. A study is compared only with an earlier study that asked the same question version in the same language, so a change of default language starts a new comparison. Look at the "compared engines" note: if an engine joined, left or failed in one of the weeks, the comparison uses only the engines that answered in both.

A useful rhythm:

• Weekly: scan New and Rising, and check any item you acted on.
• Monthly: review the full backlog and close items that are triaged as not actionable and have not recurred.
• After a large site or product change: look at the next few studies for the objection you targeted, and remember that nothing here guarantees the objection will fall.

Two related reads: AI brand sentiment: read the evidence behind a positive score and Is your brand known for the attributes you want to own?. Ordinary-answer tone is read through Sentiment trends, and Attributes is a separate weekly study with its own groups.

Common mistakes and what this cannot tell you

• Reading prominence as public opinion. The question asks for downsides. A high score means engines reached for it first, not that buyers agree.
• Ignoring the "also raised about" count. A category-wide objection may need an explanation, not a fix.
• Treating a linked source as verification. A source shows where the engine placed its citation. It does not confirm the page supports the claim.
• Acting on one engine's single answer. Look at raised-by and at whether the objection repeats over studies.
• Editing pages to game the score. These are observations of answers. Change pages to make the facts clear for buyers, and treat any effect on later studies as something to test.
• Expecting alerts or group edits. The Objections screen has no alerts or emails when an objection appears or rises, and you cannot edit, merge, rename or hide groups. Studies are not run per country or city, and no persona is used.
• Assuming past weeks exist. There is no backfill of weeks before the feature and no re-running of a stored study.

Frequently asked questions

Do objections mean AI thinks my product is bad?

No. The question is worded to ask for downsides, so engines will usually list some for any brand. The study tells you which objections engines put first when asked, and how that compares with competitors and earlier studies. Ordinary answers to your tracked prompts are read through Sentiment trends and Brand reasons instead.

What is the difference between objections and brand reasons?

Objections come from a weekly study that asks each engine directly for the downsides of each brand and groups them by meaning. Brand reasons counts the reasons that engines volunteer in answers to your tracked prompts, using thirteen fixed labels. They use different data and neither changes the other.

Should I fix every objection AI raises?

No. First check whether the quote is true, outdated or unclear, and whether competitors share it. Some objections are trade-offs to explain, some are category-wide, and some are not actionable. Fix what your evidence supports and record the rest.

Can I edit or hide an objection group I disagree with?

Not in the product. Groups cannot be edited, merged, renamed or hidden. The grouping is an AI model's judgement, so use the quotes to decide whether a group is drawn sensibly, and note any disagreement in your own backlog.

How do I know a change to my site made an objection go away?

You can watch whether the objection is New, Rising, Falling, Steady or no longer raised in later studies. That is an observation, and studies can shift for reasons unrelated to your change, such as engines joining or failing. Treat the result as evidence to weigh, not proof of cause.

Who on my team can see and run this?

Any current project member can read the screen, including viewers. Only owners and editors can press Run now. Availability depends on your plan; see plans and limits.

Next step

Open the Objections tab under Brand perception and read the quote behind your top New or Rising group before you assign anything. The Brand perception feature page explains how objections, reasons and sentiment fit together, and a study asks about your brand and each active tracked competitor, so the competitors docs show how to manage that list. If you are writing a migration or switching guide, answer switching objections directly once you know which ones engines raise.

Open Objections in DiscoveredBy