Yes. On head-to-head prompts, AI engines cite vendor-owned comparison pages all the time. In a study of 6,363 "X vs Y" software prompts, ChatGPT cited at least one of the two brands' own websites in 96.8% of answers, and Claude did so in 45.6%. But being cited is not the same as being recommended. In Google AI Overviews, 69% of cited self-promoting "best" listicles came from brands the answer then left out. The pages that get both tend to be specific, built around a table, and honest about where the competitor wins.
That gap is the whole story if you own your company's "vs" and "alternatives" pages. Your comparison content is probably being read already. The real question is whether the engine quotes it, believes it, and ends the answer with your name. Below: which prompts pull from your own site, what the evidence says about fairness (and where it stops), what a quotable page contains, and how to measure the result.
Which comparison prompts cite your own site
Not all comparison prompts behave the same way. Buyers ask consideration-stage questions in three shapes: "Asana vs Trello", "alternatives to Trello" and "best project management software". The engines treat your domain very differently in each.
Head-to-head "X vs Y" prompts: your site is usually in the sources
The clearest data comes from Minuttia's AI search visibility study, published October 13, 2026, with data by Convex Works. It ran 6,363 two-way software comparison prompts and recorded how often each engine cited the brands' own websites:
| Engine | Answers citing at least one brand's own site | Answers citing both brands' sites |
|---|---|---|
| ChatGPT | 96.8% | 77.2% |
| Perplexity | 90.5% | 49.9% |
| Google AI Overviews | 84.1% | 37.3% |
| Gemini | 55.7% | 13.7% |
| Claude | 45.6% | 11.3% |
Read the ChatGPT row as a PMM. In nearly every head-to-head answer, ChatGPT pulled at least one vendor's site into its sources, and in more than three out of four it pulled both. So when a buyer asks ChatGPT to compare you with your closest rival, it is very likely reading your domain. What's still open is which page it reads, and whether that page was written to answer the question.
The spread between engines matters just as much. Gemini and Claude cite a brand site on only about half of vs answers, and cite both sides only 13.7% and 11.3% of the time. Part of that comes from the method. Each prompt ran once. ChatGPT and Gemini were measured in the consumer apps, while Claude and Perplexity ran through APIs on lighter models, and the authors note that Claude didn't always search. Even with those caveats, "AI search" is not one surface. A page that does well on ChatGPT can be close to invisible on Claude.
The authors add their own caution: "the study doesn't specifically measure the effectiveness of comparison guides." It shows that brand sites get cited on vs prompts. It does not show that publishing a vs page causes the citation.
"Alternatives to X" prompts: you're one name in a crowd
Change the prompt to "alternatives to [competitor]" and the picture reverses. In the same study, only 3%–29% of brand mentions on alternatives searches came with a citation to that brand's official website, depending on the engine. ChatGPT had the highest rate and was still under 30%. It also named about 8.9 brands per alternatives answer on average.
In practice, an alternatives page on your site competes for attention in an answer that lists around nine names, and most of those names appear without a link to the vendor's own site. The engine is building that list mostly from other sources.
"Best X" prompts: owned pages rarely win
The Hoth looked at ChatGPT citations in the US in August 2026, using Ahrefs Brand Radar across four brands. The brand's own domain got 69.6% of citations on brand-name questions but only 1.3% on questions containing "best." The sample is small and the author sells listicle placements, so treat it as directional. B2B data points the same way. DerivateX ran 40 B2B SaaS categories through ChatGPT with web search: only 11.6% of recommendations cited the tool's own site, and 25 of the 40 categories had no own-site citations at all.
Academic work adds a wrinkle. A September 2025 paper by Chen and colleagues found that for US software queries, GPT search drew 72.7% of its sources from earned media and 26.7% from brand sites. Google's split was 45.4% earned and 43.7% brand. Comparison-style "Consideration" queries leaned even harder toward earned content. That study tested gpt-4o-search-preview on consumer categories in 2025, so rather than contradicting Minuttia's 2026 SaaS numbers, it shows how much the owned share depends on engine, year and prompt type.
If you want to see which domains ChatGPT has been citing in 2026 and how Reddit's share has shifted, this breakdown of ChatGPT citation sources in 2026 tracks where those earned-media slots went.
For a PMM, the ranking is practical. The vs page is your strongest owned surface in AI answers. The alternatives page is weaker. The self-ranked "best" listicle is weakest. Which leads to the next question: if you are getting cited, are you getting chosen?
Cited is not the same as recommended
Most citation reports mix up two things: whether the engine linked to your page, and whether the answer named you as the product to buy. The 2026 data on self-promotional content shows how far apart those can be.
Lily Ray tracked 100 B2B "best [category] software" queries in Google AI Overviews at three checkpoints between April and June 2026. A self-promoting listicle is a page where a vendor ranks itself first. Across 323 citations of those pages, the vendor was left out of the actual recommendation 224 times, or 69%. In 74 of the 100 prompts, a self-promoter's list was cited and that brand was not recommended. On "best help desk software", Pylon's and Crisp's pages were both cited, and the answer recommended Zendesk, Freshdesk and Help Scout. Ray's reading is that recommendations are "anchored to how much of the web is talking about you and recommending you."
Ahrefs ran the closest thing to a controlled test. It published 34 self-promotional pages on five domains and tracked 9,886 answers across ChatGPT, Gemini, Perplexity and Copilot from February to May 2026. For a new brand, its Evolve conference, 82% of 72 newly won mentions came in answers that cited the pages. For an established brand, Ahrefs itself, only 6% did; the rest came from third-party content. And 43% of answers citing a conference-promoting page never mentioned Evolve at all. Mateusz Makosiewicz summed it up: "The page got the citation, but the competitor got the recommendation."
Two more signals point the same way, especially on ChatGPT. Peec analysed 232,000 citations from software-review listicles between December 2025 and February 2026. Self-promotional "best of" lists made up 3.6% of ChatGPT's citations, compared with 10.3% on Google AI Mode and 10.4% on Perplexity. Seer Interactive saw ChatGPT's listicle citations fall about 30% from December to January, while total citations fell 22.7%. On the same prompts, AI Overviews stayed flat or rose.
What this evidence does and doesn't prove
Here is the honest caveat. No study we found has tested balanced and one-sided comparison pages head to head and counted citations. Most of the data above covers "best X" listicles, not "X vs Y" pages. The case for fairness is an inference from three findings:
- Self-crowning pages get read and then overruled.
- ChatGPT appears to filter them harder than other engines do.
- Engines clearly recommend competitors off your page when they judge them a better fit.
The other side deserves a fair hearing too. Glen Allsopp's Ahrefs analysis of "best" lists concluded that SaaS companies should keep publishing them, because the data showed no clear downside. Peec saw no algorithmic correction over 12 weeks. Grow and Convert says it sees no evidence of a Google crackdown across its clients. Both sides can be right, because they measure different things. Self-ranking pages clearly still earn citations. The 2026 data from Ray and Ahrefs suggests the recommendation often goes to someone else, and for a PMM the recommendation is the metric that pays.
If engines overrule a page that crowns itself, the useful question is what a page needs so the engine quotes it instead.
Why fairness gives the engine more to quote
Start with how AI comparison answers are built. Minuttia sorted the evaluation attributes engines use in vs answers. Features came up most often, followed by target fit and platform scope. Behind them came integrations, ease of use, pricing, automation, analytics, AI capabilities, scalability and security. Platform scope was the most common opening attribute, first in 48.9% of answers, ahead of target fit (13.3%) and features (10.7%).
"Target fit" is the engine's version of "who is this for?" A comparison page that only argues why you win says nothing about who the other product suits. A page with an explicit "Choose [competitor] if…" section gives the engine a ready-made, attributable answer to one of the questions it builds the response around. That's an inference, not a measured effect, but it's a well-grounded one.
Google's own guidance reads almost like a spec for this. Its Write high quality reviews documentation asks reviewers to:
- "Discuss the benefits and drawbacks of something, based on your own original research"
- "Explain what sets something apart from its competitors"
- "Cover comparable things to consider, or explain which might be best for certain uses or circumstances"
When you call something the best, it asks you to explain why, with first-hand evidence. AI Overviews and AI Mode run on Google's index and ranking, so this is the closest an engine operator has come to telling you to be fair.
Specific details matter as much as balance. The original GEO paper (Aggarwal et al., KDD 2024) reported that optimisation can "boost visibility by up to 40% in generative engine responses." Adding statistics and adding quotations were among the strongest methods. On a comparison page, the statistics are your prices, plan limits, seat counts and dates: concrete figures an engine can lift without paraphrasing. The same pattern shows up in why original research earns outsized AI citations, and a vs page that includes a benchmark you ran yourself gets both benefits.
Anatomy of a vs page that gets quoted
Put those threads together and a quotable comparison page has five parts. None of them is exotic. What separates good pages is how strictly each part is done.
1. An answer-first verdict
Open with two to four sentences that answer "which should I choose?" directly, including when the competitor is the better pick. Geoptimizer's site audit targets an answer-first intro of 40 to 80 words. That's long enough to carry the verdict and short enough to be lifted whole. If a reader has to scroll to find the verdict, an engine retrieving passages may never reach it.
2. A server-rendered HTML table with dated numbers
Of the ChatGPT-cited pages in DerivateX's B2B study, 68% included a comparison table, 56% included an FAQ, and all of them used numbered or bulleted lists. That's descriptive, with no control group, but it fits how engines handle structured content. Two details decide whether your table helps:
- The table has to exist in the HTML. In December 2024, Vercel and MERJ reported that "none of the major AI crawlers currently render JavaScript," and advised server-side rendering for critical content. Interactive comparison widgets, pricing toggles and client-rendered tabs can send an empty table to retrieval bots. Crawler behaviour may have changed since then, so check your own page.
- Each row has to make sense on its own. Chunkers often split tables by row, so a row that reads "✓ | ✗" means nothing once it's cut off from its header. Write each row so it names the attribute and gives both products' values in plain words. Our guide to how AI engines chunk a page for passage retrieval covers the mechanics.
Date any number that can change. "Starts at $12/month (as of September 2026)" is a claim an engine can quote and a reader can check.
3. A "Choose [competitor] if…" section
Give it its own heading and make the conditions specific: team size, a must-have feature, a budget threshold, an existing stack. This supplies the target-fit attribute directly. It also lets you define the competitor's segment yourself instead of leaving the engine to define it for you.
4. Sourced claims
Link to the competitor's pricing page or docs for every claim you make about them. Seer's Nick Haigler put it plainly: "Backing your claims up with verifiable information will provide those trust signals that AI models love when using external sources." Sourcing also speeds up legal review, which we come back to below.
5. An FAQ in the buyer's words
Use questions like "Is [competitor] cheaper than [you]?", "Can I migrate from [competitor]?" and "Does [you] have [feature]?" Each answer is a self-contained passage aimed at a prompt a buyer actually types.
Six vendor pages that concede well
The pattern is easier to see in real pages. All of these are vendor-owned, and each one states, in plain words, where the competitor is the better choice.
Help Scout vs Zendesk (August 2026) pairs a channel-by-channel pricing table with an eight-category summary table. Then it says: "If phone support or community forums are must-haves for your business, Zendesk will be the better option," and "If you have more than 500 employees and need enterprise-grade features, choose Zendesk." It also notes: "Zendesk does offer a skill-based routing feature that you won't find in Help Scout." These are target-fit answers an engine can lift word for word. For what it's worth, Help Scout was among the brands AI Overviews recommended for "best help desk software" in Ray's data. That's a correlation, not proof this page caused it.
PostHog vs Amplitude (January 2026) has a feature table for each product area and two H2s built for this purpose: "When to choose PostHog vs Amplitude" and "Recommendations by team type." For marketing and growth teams, it says "Amplitude is probably the right choice," because "Amplitude's UI and toolset is more geared towards marketing use cases than PostHog." It even mentions that Amplitude offers a startup scholarship.
Plausible vs Google Analytics 4 uses an eight-row table covering cookie banners, personal data, ad-blockers, modelled data, retention, setup, EU hosting and open source. It includes a "What GA4 does better" section. That section says GA4 with BigQuery "has more depth," and that for heavy Google Ads users, "the tight integration between GA4 and Google's ad products is hard to replicate elsewhere." Then it reframes: that depth "comes with overhead that is rarely justified" for most site owners. Concede, then reframe. The concession is what makes the reframe believable.
Zapier vs Make (Zapier's blog, September 2026) has an eight-row table and concedes on price with specific figures: "Make's entry-level plan does start lower ($12/month vs. Zapier's $19.99/month)." It adds that "Make can work well for a small, technical team that wants deep control over branching logic."
Supabase vs Firebase has seven comparison tables. It grants that "Firebase offers an excellent developer experience for prototypes and simple applications," and that it "excels at real‑time synchronization and has a generous free tier."
A useful middle case is Attio's "Best Salesforce CRM alternatives in 2026". Attio ranks itself first. But it says Salesforce "built the category, and for a certain kind of company, that's still the right call." It also lists its own limitation: it's newer to the market than Salesforce or HubSpot, so its app ecosystem is smaller. There's no table. It's a self-ranked page with honest limits, which is a reasonable compromise if your team won't publish a page that doesn't put you first.
Notice what these pages have in common. None of them is neutral; each still argues for its own product. They just make the case to a defined buyer and say plainly who that buyer isn't.
The PMM checklist: legal, freshness and plumbing
A good draft still has to pass legal review and stay accurate for months. These are the checks to run before publishing, and again whenever a competitor changes plans.
- Legal: truthful comparison is encouraged. The FTC's 1979 policy statement on comparative advertising says the Commission "encourages the naming of, or reference to competitors." It requires "clarity, and, if necessary, disclosure to avoid deception," and holds that "disparaging advertising is permissible so long as it is truthful and not deceptive." Dated, sourced, specific claims, including concessions, are the easiest to defend. This is context, not legal advice; your counsel makes the call.
- Accuracy, because engines repeat you. In January 2026, Lily Ray published an AI-generated article about a Google core update that never happened. Within 24 hours AI Overviews served it back as fact, with her site as the only source. For niche pairs, your vs page may be the only source the engine finds. Whatever it says about the competitor, accurate or out of date, can become the answer.
- Fresh, but for real. In Ahrefs' study of 1,100 cited, dated "best" lists, 79.1% had last been updated in 2025, and 26% in the two months before the study was published. But Ray flagged "artificial refreshing" as a pattern on SaaS sites that lost visibility: adding the year to a title without substantive updates. Re-check competitor pricing first, then change the date.
- No self-serving review markup. Under Google's review snippet guidelines, self-serving reviews on LocalBusiness or Organization pages aren't eligible for star review features. The guidelines also say: "Don't aggregate reviews or ratings from other websites." Ray listed AggregateRating schema on ineligible pages among the patterns she saw on demoted SaaS sites.
- The table has to reach the bot. Fetch the page the way retrieval bots do. Our AI crawler access checker reads robots.txt and llms.txt, then live-fetches the page to catch CDN blocks that robots.txt can't show. If the table isn't in what comes back, it isn't in the answer.
One more data point if your blog has self-ranked pages. Ray reported Sistrix visibility drops of 29% to 49% in January 2026 on anonymised SaaS sites with many self-promotional listicles. She says herself that the evidence is correlational and from a selected sample, and Grow and Convert disputes the causal link. Either way, her summary is the right stance: "it works, until it doesn't."
How to tell whether the page worked
Publishing is the easy part. Knowing whether it changed anything is harder, for three reasons.
Citations flicker. In Ahrefs' experiment, about a quarter of citations appeared once for a given query and engine and never again. Suppose you type "[us] vs [them]" into ChatGPT on launch day and see your page cited. You've learned very little, and you'd learn just as little if it were missing.
Engines diverge. ChatGPT cited a brand site on 96.8% of vs answers in the Minuttia data, and Claude on 45.6%. That gap means you need a read for each engine. Grok is a further blank: none of the studies above covered its behaviour on comparison prompts, so you'll need your own data there.
Citation is only half the result. You also need to know whether the answer names you, how early, and how favourably.
A workable measurement setup looks like this:
- Track consideration-stage prompts directly. That means "[you] vs [competitor]" for each vs page you own, "[competitor] alternatives", and "best [category] for [the segment you claim]".
- Read citation and recommendation separately. A high citation rate with low prominence or neutral sentiment is the cited-but-not-recommended pattern showing up in your own data.
- Use a rolling window, not a snapshot, so one flicker doesn't look like a trend.
- Mark the launch or rewrite on your timeline, along with engine model changes, so a drop isn't wrongly blamed on the page.
Geoptimizer was built for exactly this kind of measurement. It runs your tracked prompts on ChatGPT, Gemini, Claude and Grok with web search on. For each engine it reports mention rate, citation rate, prominence and sentiment over a 7-day rolling window. Verbatim answer excerpts show where in each answer your brand is first named. Fixes and model changes are marked on the chart, so you can see whether the vs-page rewrite moved anything.
Expect a ceiling, though. In Ray's data, recommended brands generally had far more referring domains and AI mentions, and the Chen paper describes a "big brand bias" that hurts niche players. A fair vs page won't make up for a thin off-site footprint. Pair it with work on getting into the third-party "best tools" lists AI engines cite.
FAQ
Do AI engines cite vendor-owned comparison pages?
Yes, especially on head-to-head prompts. Minuttia's October 2026 study covered 6,363 "X vs Y" software prompts. The share of answers citing at least one of the two brands' own sites was 96.8% on ChatGPT, 90.5% on Perplexity, 84.1% on Google AI Overviews, 55.7% on Gemini and 45.6% on Claude. On "alternatives" and "best" prompts, brand sites are cited far less.
Should my comparison page admit where the competitor is better?
The evidence suggests it helps, though no study has tested it directly. Self-crowning pages get cited and then overruled. In Lily Ray's AI Overviews data, 69% of cited self-promoting listicles came from brands the answer didn't recommend. A "Choose [competitor] if…" section also supplies the target-fit attribute that AI comparison answers are built around, and it matches Google's review guidance.
Are "alternatives to [competitor]" pages worth publishing?
They're a weaker owned surface than vs pages. Only 3%–29% of brand mentions on alternatives prompts came with a citation to the brand's own site, and ChatGPT named about 8.9 brands per answer. Publish them if the content is genuinely useful, but expect third-party sources to shape most alternatives answers.
Will AI crawlers read a comparison table built in JavaScript tabs?
Possibly not. Vercel and MERJ reported in December 2024 that none of the major AI crawlers rendered JavaScript. Server-render the table, then fetch the page as a retrieval bot to confirm the rows are in the HTML.
How often should I update a vs page?
Update it whenever a competitor's pricing, plans or key features change, and date the figures you show. Change the page's date only when the content actually changed. Lily Ray flagged year-bumping without substantive updates as a pattern on SaaS sites that lost visibility.
Start with the prompts your buyers compare
AI engines are probably already reading your vs pages. What you don't know yet is whether they quote them, on which engines, and whether the answer ends with your name or your competitor's. Write the page fairly, specifically and around a table. Then measure the result instead of guessing. You can track your "vs" and "alternatives" prompts on all four engines with Geoptimizer's free plan, no credit card needed. Compare what each Geoptimizer plan includes and start your chart.