Key takeaways
- Comparison prompts are the riskiest prompts for your brand. In Promptwatch Data (Sep 28 to Oct 4, 2026), 49.1% of ChatGPT responses to competitor comparison prompts cite 10 or more domains, so the answer is stitched together from many sources, including ones you don't control.
- Most comparison errors trace back to a small set of pages: your own product and pricing pages, a competitor's "vs" page, and third-party review or roundup sites.
- Fixing it is a loop: collect the actual "X vs Y" answers, mark each claim right or wrong, find the source page, correct or replace it, then re-check.
- Publish your own balanced comparison pages. Comparisons are only 5.9% of ChatGPT citations overall but 7.9% on commercial prompts, and product and landing pages carry far more weight.
- Since ChatGPT started using the
site:operator at scale on Aug 8, 2026, your own domain is a direct retrieval surface. Thin or unindexed pages cost you answers.
Why 'X vs Y' answers go wrong
A buyer types "YourBrand vs Competitor" into ChatGPT or Gemini. The answer comes back confident, tidy, and partly wrong. Maybe it quotes a pricing tier you retired a year ago. Maybe it says you lack a feature you shipped in spring. Maybe it just flips the verdict so the competitor looks like the safe choice.
This happens for a few boring reasons, and none of them involve the model "having it in for you."
- The sources disagree. The model pulls from many pages, and some are stale.
- The model is built to be agreeable. A prompt like "why is X better than Y" tends to get an answer that accepts the premise. One write-up on prompt bias points out that models are optimised to be helpful and to go along with your framing, which is the opposite of what you want when a buyer has already half-decided.
- Errors are common in general. A BBC study, as summarised by Josh Bersin in October 2025, found 45% of AI answers it tested contained errors. That figure is secondhand and not about comparisons specifically, but it sets a realistic baseline: don't assume the answer about you is right.
- Competitors write the "vs" pages. Their version of the story is often the most detailed one on the web.
The cost is real. Buyers ask AI to compare providers before they talk to sales, as the Shoden Global FAQ on comparison content puts it. If the comparison is wrong, you lose the deal in a conversation you never saw.
The common types of comparison errors
Before fixing anything, name what you are fixing. I find it helps to sort errors into five buckets.
| Error type | What it looks like | Usual source |
|---|---|---|
| Outdated pricing | Old plan names, wrong price points, a free tier that no longer exists | Old review sites, cached roundups, stale pricing pages |
| Missing or wrong features | "Doesn't support X" when it does | Competitor "vs" pages, old docs, old reviews |
| Wrong positioning | Labeled as an "enterprise tool" when you target small teams, or the reverse | Third-party listicles, directory descriptions |
| Wrong verdict | Competitor recommended by default, no reasoning given | Prompt framing, thin or absent comparison content from you |
| Invented details | Claims with no source at all | Model gap-filling from training data |
The last one is the hardest to fix, because there is no page to edit. The first four almost always have a source you can find.
Why comparison prompts are different from other prompts
Promptwatch Data has a useful split here. In its report on how many unique domains ChatGPT cites per response (Sep 28 to Oct 4, 2026, responses with at least one citation):
| Prompt type | Cite 10+ domains | Cite 5 or fewer | Most common pattern |
|---|---|---|---|
| Competitor comparison ("Asana vs Monday.com") | 49.1% | not reported | Only 0.35% cite a single domain |
| Organic (no brand named) | 52.5% | not reported | 4.5% cite 3 or fewer |
| Brand-specific ("is Asana good for small teams") | 21.3% | 40.7% | Most common count is 5 domains (14.9%) |
Two things stand out. First, ChatGPT almost never settles a comparison with one site. It casts a wide net, so you can't fix the answer by fixing one page. Second, brand-specific prompts draw from a short list. When someone asks about you alone, the handful of review sites and forums on that source list shape the description. Those are the pages to audit first.
The report's own advice is worth repeating: expect your competitor's site to be cited next to yours, and spell out your differences clearly on your own pages so they are easy to quote.
Step 1: Capture what AI actually says
You can't fix what you haven't read. Start with a manual pass, then automate.
Build a list of comparison prompts
Pull these from three places:
- Your own sales calls and win/loss notes ("we keep getting compared to...")
- Search Console queries containing "vs", "alternative", or "compared to"
- Your top three to five named competitors, in both orders ("A vs B" and "B vs A"), because the order can change the answer
Add variants with the buyer's context: "for small teams", "for agencies", "on a budget", "for enterprise". Persona-style phrasing matters because the answer shifts with it.
Run them on the real interfaces
Answers in the actual ChatGPT, Gemini, Perplexity, Copilot, and Google AI Overviews interfaces can differ from what an API returns. If you test through a developer API only, you are looking at what the model believes from training data, not what a buyer sees with live search on.
Save every answer and every cited URL. A spreadsheet works for a first pass. After about thirty prompts you will want a tool.
Tools that can help
A monitoring platform saves you from doing this by hand across models and countries. Promptwatch tracks the real interfaces of ChatGPT, Gemini, Perplexity, Google AI Overviews, and others, and its Brand Narrative feature pulls out the individual claims AI makes about you and competitors ("affordable for small teams", "lacks a mobile app") so you can review claims instead of whole paragraphs.
Other trackers exist and are worth a look if your needs are narrower. Peec AI, Otterly.AI, and Profound all monitor brand mentions in AI answers, though they stop closer to reporting than fixing.
Otterly.AI
Profound
For a wider list, the GEO software directory at bestgeosoftware.com compares options by feature.
Step 2: Grade each claim
This step is tedious and it is the one people skip. For each saved answer, break it into claims and mark each one:
- Accurate
- Inaccurate
- Outdated (was true, isn't now)
- Unverifiable (no source, can't tell)
Then tally by claim, not by answer. You will usually find that three or four claims account for most of the damage, repeated across prompts and models. "Lacks integration X" showing up in eleven answers is one problem to fix, not eleven.
In Promptwatch, Fact Check does this on the Business plan and up: you mark a claim Accurate or Inaccurate, and it reads the pages crawled from your site to find the ones that state the claim or its opposite. That tells you whether the wrong claim is coming from your own site, which happens more often than teams expect. A spreadsheet and a site search can do the same job slowly.
Step 3: Trace each wrong claim to its source
Every wrong claim lives somewhere. Look in this order:
- Your own pages. Old pricing in a blog post, a retired feature on a product page, a comparison PDF nobody updated. Fix these first because they are fully in your control.
- The cited sources in the answer. Open every URL the answer cites. Look for the sentence that matches the wrong claim.
- Competitor "vs" pages. If the competitor has published "Them vs You," read it as a buyer would. It is probably the most detailed page on the topic.
- Review and directory sites. G2, Capterra, Trustpilot, and similar sites often hold the profile text ChatGPT quotes for brand-specific prompts.
- Roundups and listicles. Promptwatch Data shows listicles are 13.6% of ChatGPT citations overall, and 15.0% of commercial ones. Your absence or misdescription in a "best X tools" post matters.
Check how AI reaches your pages
If your correct page exists but never gets cited, check whether AI crawlers can read it. Robots.txt blocks, JavaScript rendering problems, and server errors are common culprits. Promptwatch's AI crawler logs show which bots hit which pages and what errors they got, which separates an access problem from a content problem. The free Agentic Readiness Check gives a quick first read on which AI crawlers can reach your site.
Step 4: Fix the pages, in priority order
Promptwatch Data on content types ChatGPT cites (Sep 28 to Oct 4, 2026) tells you where effort pays off.
| Content type | Share of all ChatGPT citations | Share on commercial prompts | Share on branded prompts |
|---|---|---|---|
| Product pages | 22.4% | 26.0% | 23.3% |
| Landing pages | 16.8% | 20.6% | 22.8% |
| Listicles | 13.6% | 15.0% | not reported |
| Comparisons | 5.9% | 7.9% | not reported |
| Reviews | 1.7% | not reported | 5.6% |
| FAQ pages | 3.3% | not reported | 4.8% |
Product and landing pages together are 39% of citations. So the unglamorous fix, getting specs, prices, and positioning right and easy to quote, beats most clever tactics. The report says ChatGPT quotes those pages directly.
Fix product and pricing pages
- State the current price and plan names in plain text, not only inside images or JavaScript widgets.
- List what each plan includes and who it is for.
- Put a "last updated" date on pricing and feature pages.
- Remove or redirect retired plan pages. Old URLs keep getting cited long after you forget them.
Publish your own comparison pages
If you don't write "You vs Competitor," someone else will. Make yours worth citing:
- Be fair. A page that only flatters you reads as marketing and gets quoted less. Say where the competitor is better. It builds trust with buyers and gives AI something balanced to reuse.
- Say who each option is best for, in one or two sentences, near the top.
- Use a comparison table with specific, checkable rows: pricing, key features, integrations, limits.
- Add a short FAQ for the questions buyers actually ask.
- Show proof: screenshots, customer numbers, dates, links to docs.
The Shoden Global FAQ lists the mistakes that make comparison content hard to reuse: vague claims, biased writing, no table, no clear recommendation, weak structure, missing FAQs, no proof. That list is generic but accurate.
Watch your phrasing too. A style note that circulated in April 2026 from Ruben Hassid points out that "It's not X, it's Y" constructions are a giveaway of AI-written text, and Barron's counted them quadrupling in Fortune 500 filings between 2023 and 2025. Comparison pages full of contrastive filler read as machine-made, and they say little. "Competitor A has no native Salesforce sync; we do" is more useful than a sentence built around a rhetorical flip.
Cover more angles, not just one 'vs' page
On Aug 8, 2026, ChatGPT began using the site: operator at scale. Per Promptwatch Data, the share of fanout queries using it jumped from 0.37% to 16.8% in a day, and average fanout queries per response rose from about 1.08 to about 1.83. The extra searches are added on top of the generic ones.
For you, that means ChatGPT may search your domain directly for your pricing, your alternatives, and your comparisons. A site: query only surfaces pages search engines have indexed. Thin category pages, broken internal search, and unindexed content now cost you answers. Treat comparisons, pricing, and alternatives as separate retrievals and give each its own indexed page.
Fix third-party descriptions
This is slower and partly out of your hands, but worth doing:
- Claim and correct your profiles on review and directory sites. Update descriptions, pricing, and screenshots.
- Ask happy customers for fresh reviews on the sites your buyers trust. Branded prompts cite reviews at 5.6%, more than three times the overall rate.
- Contact authors of roundups that misdescribe you. Send a short, factual correction with a link to your docs. Don't argue.
- Pitch inclusion in the "best of" lists that ChatGPT already cites for your category.
A word on Reddit. It is tempting to go correct misinformation in threads. But per Promptwatch Data, reddit.com's share of ChatGPT citations fell from about 4% to 0.5% on Aug 14, 2026, and ChatGPT cites social platforms for only 0.46% of citations (Sep 28 to Oct 4, 2026) versus 12.2% for Google AI Overviews. See the Reddit citation drop report and the social media citations report. Reddit still matters for AI Overviews, so it is a channel worth watching, not a first fix for ChatGPT.
Step 5: Re-check, and keep re-checking
Fixes don't show up overnight. AI answers change when a page is re-crawled and re-indexed, and when the model's retrieval shifts. Content also ages: Promptwatch's content half-life data shows news losing half its cited pages within 36 days, while landing pages and docs stay cited longest. Put your key claims on durable page types.
A simple cadence:
- Weekly: rerun your top ten comparison prompts and diff the claims.
- After every pricing or feature change: update the pages, then rerun the affected prompts.
- Monthly: review which domains appear on the source list for brand-specific prompts.
- After a platform shift: the
site:change was overnight and platform-wide, so continuous monitoring beats one-off audits.
Page-level tracking helps close the loop. Promptwatch's Page Tracker shows which prompts and responses cite a given URL, so you can see whether the corrected page replaced the stale one. You can also use its Content Agent on the Business plan to draft comparison pages grounded in the prompts and competitor pages AI already cites, with a review step before anything is published. Check pricing before you commit, and the 7-day trial or the free AI visibility report is a low-risk way to see your own baseline.
Prompt framing and why you should test it
Buyers don't phrase prompts neutrally. "Why is Competitor better than you" and "Compare you and Competitor" produce different answers, partly because models lean toward agreeing with the premise. Test both framings.
A practical set:
- Neutral: "Compare A and B for a 20-person team."
- Leading toward you: "Why choose A over B?"
- Leading against you: "What are the downsides of A compared to B?"
- Alternatives: "Best alternatives to B."
- Verdict: "Which should I buy, A or B?"
The leading-against prompt is the one to study. It shows which weaknesses AI already associates with you. Some will be fair. Fix the unfair ones with better pages, and take the fair ones to your product team. I would rather learn that from a prompt log than from a lost deal.
What you can't fix
Be realistic. You can't edit a model's training data. Some errors come from what the model absorbed long ago, with no clean source to correct. For those, the lever is volume and consistency: more current, correct, indexed pages saying the same true thing, across your site and the third-party sites that matter. It works slowly and not every time. A strong source list for brand-specific prompts helps most, since those responses draw from a short list of sites.
A quick checklist
- List 20 to 30 comparison prompts, in both orders, with persona variants
- Capture answers from real interfaces, not just APIs
- Grade claims as accurate, inaccurate, outdated, or unverifiable
- Trace each wrong claim to a source URL
- Update your product, pricing, and feature pages first
- Publish fair, tabled comparison pages for your top competitors
- Correct review-site profiles and roundups
- Confirm AI crawlers can reach the corrected pages
- Rerun the prompts weekly and keep a diff
If you want a broader view of tools in this space, the AI rank tracking directory at ai-rank-tools.com is a good place to compare trackers, and Promptwatch's comparison of GEO platforms lays out feature differences side by side. Start with the three claims that show up most often, fix their sources, and rerun the prompts next week.