Defensive GEO: What Does ChatGPT Say When Buyers Ask "Why Shouldn't I Buy Your Product?"

Negative answers are rare, but they land at a bad moment. In BrightEdge's March 2026 analysis, ChatGPT showed negative sentiment in about 1.6% of brand mentions and Google's AI Overviews in about 2.3%. Yet 19.4% of ChatGPT's negative sentiment came up at the consideration stage, against 1.5% for Google. BrightEdge adds that ChatGPT's criticism "more frequently reflects product reviews, forums, and social discussions such as Reddit."

Most GEO guides are about offense: getting named in "best X" answers. Defense gets far less attention, though one sentence can be costly. Wolf River Electric, a Minnesota solar installer, alleges in a lawsuit that a Google AI Overview falsely said the state's attorney general had sued it, and that a customer cancelled a $150,000 contract over it. Those are claims in an ongoing case, not court findings. They still show the stakes.

This playbook covers the defensive prompts to run, how to classify and trace what comes back, the counter-narrative page, the rules for replying on Reddit and review sites, and when the honest fix is the product. For the weekly version, see our guide to monitoring brand mentions in AI-generated responses. For one wrong fact, use the quicker report, correct and recheck routine.

What Defensive GEO Is, and Why Offense Isn't Enough

Offensive GEO asks, "Does the AI name us?" Defensive GEO asks, "When the AI names us, does it hand the buyer a reason to leave?" The first grows the shortlist. The second protects the deals already on it.

The prompts differ too. Offense runs unbranded category prompts. Defense runs branded prompts with a negative frame. A brand can do well on the first set and still lose sales on the second, because the buyer who asks "What's wrong with Plannora?" is close to a decision.

Where negative claims come from

Search-grounded answers lean on pages anyone can publish, and forums rank high among them. Reddit held 18.5% of the citation share among the 50 domains Google's AI Overviews cited most in Ahrefs' September 2026 count, and 21.6% among Perplexity's top 50. OpenAI has had access to Reddit's Data API since a May 2024 partnership that brings "real-time, structured, and unique content from Reddit" into ChatGPT. Old complaint threads rarely get an update, so a 2024 thread can keep answering a 2026 question. That makes forums the first place defensive GEO looks.

Competitors are a second source. In a December 2025 experiment, the agency Reboot published negative claims about a made-up person on 10 existing websites and watched 11 AI models. Weeks later, only Perplexity and ChatGPT cited the test sites. Perplexity repeated the claims with hedges such as "reported as," while ChatGPT "explicitly questioned the credibility of the sources." The other nine models never picked them up. A lone, uncorroborated claim got limited traction, which is why the defense is corroborated evidence.

Fast layer, slow layer: what publishing can change

It's tempting to aim for content that models "ingest," so a competitor's complaint never becomes the AI's permanent answer. Be precise here, because every answer has two layers, and you can only move one of them quickly.

OpenAI says its models learn from sources that include "public forums," and that models are "large sets of numbers, known as 'weights'" that "do not store or retain copies" of training data (OpenAI Help Center). You can't edit weights. You can make the corrected record the one search finds today and the one a future training crawl sees. And ChatGPT only searches when it judges that a question "would benefit from current information" (OpenAI), so an answer written from memory won't change because you published a page last week. The grounding layer is where defensive GEO moves first.

The Defensive GEO Prompt Set: 15 Ways Buyers Ask "Why Not You?"

A defensive GEO audit starts with prompts written the way a doubtful buyer would write them. Swap in your brand, your main rival and your buyer's situation. The examples use Plannora, a made-up project management tool, and Loopcraft, a made-up rival.

Pick the comparisons you already lose. Your sales team knows them: the deals that went to a rival, and the reason the buyer gave. A rival matrix like the one in how to benchmark AI citations against competitors shows which rival wins most often.

How to run the set

1. Use a clean session. Log out or use a temporary chat, turn memory off, and run from the same place each time.
2. Run each prompt three times per engine. One answer is an anecdote. Our post on the technical reality of tracking AI answers explains why answers drift.
3. Note whether ChatGPT searched. If it answered from memory, OpenAI lets you regenerate and "choose Try again or Search the web, when available." Comparing the two versions tells you which layer a claim lives in.
4. Save the full answer and every cited URL.
5. Pull out each negative sentence about you. Skip general caveats about the category.

The free AI Visibility Prompt Kit gives you a starter list you can rework with these negative frames.

The Objection Ledger: Classify Every Negative Claim

The Objection Ledger is the working sheet of defensive GEO: one tab, one row per negative claim, grouped into clusters of the same complaint. It turns "ChatGPT is harsh about us" into a short list with an owner on every line. Give each cluster one of five verdicts.

For each cluster, record the claim in the AI's words, the prompts that triggered it, how many answers carried it, which engines, the verdict, the main cited URL, the source type (owned, review site, Reddit or forum, rival page, news, or none) and an owner.

Rank the clusters

Sort by how many answers carry each claim, then by verdict. False and fixed-but-cited claims come first, because they're the cheapest to move. True claims go to the product team with the count attached.

The same sheet gives you the Accuracy Rate from our GEO Metrics Framework: answers with no false or outdated claim, divided by all answers that name you. Report it next to the count of negative claims, so a fall in one isn't hidden by the other.

Trace Each Negative Claim to Its Source

Every claim you want to change needs a URL behind it, and tracing is where most defensive GEO time goes. Each engine shows its sources differently.

• Perplexity puts clickable citations inside the answer, so open each one next to the sentence it supports. Our guide to brand presence in Perplexity walks through that trail in detail.
• ChatGPT shows a Sources button when it searched. OpenAI warns that "search results and citations can be incomplete, outdated, or incorrect," so open each page and find the passage.
• Google AI Overviews include links to the pages that support them. Google's own help page says AI Overviews "can and will make mistakes."
• No sources at all usually means the answer came from training data, or the model joined facts that no single page states. Wolf River's complaint alleges the second kind: that none of the linked pages said what the overview claimed.

When the source is a rival

A competitor's comparison page often sits under a losing comparison. The FTC's policy on comparative advertising says "disparaging advertising is permissible so long as it is truthful and not deceptive." So you can't stop a fair critique, and you shouldn't try. You can challenge a false one: email the rival with the wrong line, the current fact and a link to proof, and keep a record. If a false claim is doing real damage, that's a question for a lawyer.

Then publish your own fair comparison, one that admits where the rival wins, so the engine has a second view to weigh. That is how you keep a rival's framing from becoming the default answer: not by hiding it, but by making sure it's never the only page on the question.

Worked Example: Plannora's First Defensive GEO Audit

Plannora, Loopcraft and stackreview.co are made up, and so are these numbers. The arithmetic is real. Plannora ran the 15 prompts three times each in ChatGPT, with search on, and in Perplexity: 90 answers. Of those, 62 (68.9%) carried at least one negative claim, 118 claims in all, in six clusters.

Four clusters (A, B, D and F) are source problems: 81 of 118 claims, or 68.6%. Cluster C, 22 claims (18.6%), is a product gap. Cluster E, 15 claims (12.7%), is a judgment about fit.

Before any fix, 49 of the 90 answers held at least one false or outdated claim. So 41 of 90 were accurate, an Accuracy Rate of 45.6%.

The response plan, in order of size and cost:

1. A: add a dated entry for the sync bug to the counter-narrative page, link the changelog, and reply in the Reddit thread from a disclosed staff account. Ask the YouTuber to pin a correction.
2. B: send correction requests to Loopcraft and the listicle author, and state the guest rule in one sentence on the pricing page. Hidden-cost claims often start with pricing gaps, as our post on hallucination by omission explains.
3. D: ask stackreview.co to add the date chat support launched.
4. F: publish a data residency page that names every hosting region.
5. C: send the count to the product team: 22 of 118 claims.
6. E: add a "Who Plannora is for, and who it isn't" section to the counter-narrative page.

Eight weeks later, Plannora reran the same 90 runs the same way.

Source-fixable claims fell 79.0%, from 81 to 17. The true complaint barely moved, from 22 to 20, and the fit judgment didn't move at all. That's the lesson of the exercise: pages change what's false or stale, not what's true.

Is the drop real? The share of answers with a negative claim fell 31.1 points. The 95% margin for two samples of 90 at these rates is 1.96 × √(0.689 × 0.311 ÷ 90 + 0.378 × 0.622 ÷ 90) = 1.96 × 0.0707 = 13.8 points. A 31.1-point fall clears it.

Publish a Counter-Narrative Page Buyers Can Trust

The counter-narrative page is the main asset in defensive GEO. It answers the negative questions on your own site, in the open. Search-grounded engines look for a page that answers the question asked, and today most pages about "Plannora complaints" were written by someone else. The page isn't spin. If it reads like spin, buyers will treat it that way, and so will any system built to find reliable sources.

What goes on the page

• A plain title that matches the question, such as "Plannora limitations and common complaints, answered (updated September 2026)".
• One section per complaint, headed with the question in the buyer's words.
• A verdict in the first sentence: true, fixed on a stated date, or not true.
• Specifics: version numbers, dates, prices, regions and limits. "Fixed in version 3.2, June 2024" beats "we've improved sync."
• Proof links to the changelog, status page, docs or pricing.
• The true limits, stated honestly. "Plannora has no offline mode. If your team often works without internet, Loopcraft may suit you better." Admitting a real limit is what makes the rest believable.
• A visible "last updated" date, a named owner and a short change log at the bottom.

Link the page from pricing, comparison pages and the help center, so buyers and crawlers can reach it. Google's AI optimization guide favors "non-commodity" content built on first-hand knowledge. A page only you can write, with your own version history, fits that description.

Format and markup

Use real headings and short answers, the same layout as an FAQ. FAQPage schema is optional. Google stopped showing FAQ rich results on 7 May 2026, according to its Search Central changelog, and its AI guide says "structured data isn't required for generative AI search." For the schema pattern on a facts page, see how to fix incorrect brand facts in AI answers.

One page, kept current, beats twenty thin variants. Google warns that making separate pages for every phrasing "primarily to manipulate rankings or generative AI responses" violates its scaled content abuse policy.

What never to do

• No fake reviews. The FTC's 2024 rule bans fake and AI-written reviews, reviews bought for a set sentiment and undisclosed staff reviews, and allows civil penalties for knowing violations.
• No sock puppets or vote rings. Reddit bans "multiple accounts, voting services, or any automation to manipulate vote counts" in its policy on disrupting communities.
• No planted "independent" articles. Google says "seeking inauthentic 'mentions' across the web isn't as helpful as it might seem," and since May 2026 its changelog states that its spam policies also apply to generative AI responses.

Respond on Reddit and Review Sites Within Their Rules

Much of defensive GEO happens off your site. You can't delete someone's thread or review, but you can add the missing fact where buyers and engines will read it.

Reddit

Reddit's rules ask users to "participate authentically" and not to "impersonate an individual or entity in a deceptive manner." Its spam guidance asks people whose posts mostly link to their own business to "be thoughtful about the frequency of posting," and says each community's moderators decide what counts as spam there.

A reply that fits those rules:

1. Post from an account that says who you are, for example "Dana from Plannora support."
2. Answer the one complaint: what happened, the fix date and a link to the changelog or help page.
3. Read the subreddit's rules first, and ask the moderators if links to your own site are allowed.
4. Don't argue with the original poster, and never ask colleagues to upvote.

Review sites

• G2's community guidelines invite sellers "to provide their perspective and address the user's comments," but G2 doesn't "edit or remove reviews at a seller's request." The guidelines ban collecting reviews in ways that screen out negative ones, and they bar employees of the vendor and of direct competitors from reviewing.
• Trustpilot's June 2026 guidelines for businesses ask for replies that are "polite and professional." You may flag a review for reasons such as defamation or no genuine experience, but Trustpilot says it doesn't tolerate misuse of flagging, and it bans selective invitations and incentives.

The honest answer to an old review is a public reply with the fix date. The honest answer to too few reviews is to invite every customer, not only the happy ones.

Tell the AI vendors, but don't wait on them

• ChatGPT: thumbs down, then choose an issue, or use OpenAI's content report form. OpenAI's reporting page says reported domains "may be reviewed by OpenAI's Model Quality team, which may apply filters or other mitigations."
• Perplexity: the flag icon below the answer, or a support ticket, with the thread link and the answer you expected.
• Google AI Overviews: thumbs down, then "Report a problem." Google says your feedback "helps us improve AI Overviews."

None of the three promises to correct a specific answer or to reply to you. In defensive GEO, vendor feedback is a side channel: report the answer, then fix the source.

When the Complaint Is True, the Fix Is the Product

Some objections survive every correct page because they're correct. Plannora's "no offline mode" is one. No page or reply will make an honest engine stop saying it, and it shouldn't.

At this point defensive GEO becomes product research. When buyers ask AI what's wrong with you, the answer is a summary of what customers have written for years. When a true claim sits near the top:

1. Hand the count to the product team. "22 of 118 negative claims" puts the gap in buyer terms.
2. Decide to fix, price or position. Build it, change the plan rule that causes it, or narrow who you sell to.
3. Say it plainly until it ships. State the limit on the counter-narrative page and in sales material.
4. Leave a dated trail when it ships. Add a changelog entry and a "fixed" note on the counter-narrative page, then reply in the threads that raised it.

Then keep watching. Run a full defensive GEO sweep each quarter, plus a check after each launch, price change or incident. The companion guide to monitoring AI-generated responses sets up the weekly version with severity levels and an escalation path.

Where Rankbox Fits in Defensive GEO

Rankbox doesn't track AI mentions, citations or sentiment, and it doesn't monitor, correct or report what AI says about your brand. The prompt runs and the Objection Ledger stay a manual job, or one for a third-party tracker.

Rankbox helps with the pages the ledger asks for. Answer-Space Research maps the questions buyers ask ChatGPT, Perplexity and Google in your category and scores each for volume, difficulty and intent. The Citation-Ready Writer researches the live web and writes source-backed articles, such as a dated counter-narrative page or a fair comparison, which reach your site through the Rankbox API. Reddit Presence, which is rolling out, finds relevant threads and drafts replies for you to post yourself; it never posts. The Business plan is $49.50 a month with a 7-day trial: see pricing.