How ChatGPT Decides Which Software to Recommend (And Why It Keeps Naming Your Competitor)

Quick Takeaways

• ChatGPT recommends software through two separate mechanisms, training data and live retrieval, and each one rewards different things
• ChatGPT and Perplexity share only 11% of the domains they cite, so ranking in one engine says almost nothing about the others
• Listicles earn roughly 25% of AI citations, and Perplexity favors content updated within the last 30 days by 3.2x
• Review platforms lost most of their organic traffic, yet 88% of AI Overview review citations still go to just five review sites
• The fix is not a trick: get into the listicles that get cited, publish definitional pages, keep your entity facts consistent everywhere, and keep content fresh

Your competitor is not winning in ChatGPT because their product is better. They are winning because the model can describe them, and it cannot describe you.

That distinction matters more every quarter. 51% of B2B software buyers now start their research in an AI chatbot, up from 29% a year earlier, and ChatGPT alone holds 63% of that behavior. The same research found 69% of buyers chose a different vendor than they originally planned based on AI guidance. We covered the buyer's side of this shift in how B2B buyers use ChatGPT to shortlist vendors. This post is the sequel: what is happening inside the machine, why it keeps naming the same vendors, and how to audit and fix your own position.

Two engines under the hood: training data and retrieval

When ChatGPT answers "what's the best tool for X," it draws on two very different systems, and most vendors conflate them.

Training data is the model's long-term memory. It was built from a snapshot of the web, so it reflects what was widely written about you months or years ago. If your category barely existed at the last training cut, or your positioning changed since, the model's baked-in picture of you is stale or missing. You cannot edit training data directly. You can only influence the next snapshot by being described consistently, in many places, over time.

Retrieval is what happens when the model searches the live web before answering. This is where citations come from, and it behaves much more like SEO: the model runs searches, reads a handful of pages, and synthesizes an answer with sources. Retrieval is fast to influence. A page you publish this month can be cited next month.

The practical consequence: a vendor with years of consistent third-party coverage wins the training-data layer, and a vendor with fresh, well-structured, directly-answering pages wins the retrieval layer. Your competitor who keeps getting named has usually done both. Most companies have done neither deliberately.

The citation data: engines do not agree with each other

The most useful recent finding in this space is how little the engines overlap. ChatGPT and Perplexity share only 11% of the domains they cite. Read that again: 89% of the sources one engine trusts, the other ignores. There is no single "AI SEO" leaderboard. Being recommended in ChatGPT and being recommended in Perplexity are two different projects with two different source pools.

Two more numbers from the same research shape the playbook:

• Listicles get roughly 25% of AI citations. "Best X tools" and "top alternatives to Y" roundups are the single most cited content shape, because they match the literal form of the question buyers ask.
• Perplexity favors freshness hard. Content updated within the last 30 days is cited 3.2x more often. A great page from 2024 that nobody has touched is quietly aging out of the answer.

This explains the most common frustration we hear: "We rank #1 on Google for our category, but ChatGPT recommends three competitors and not us." Google rankings and AI citations are correlated, not identical. The model is not reading the SERP the way you do. It is pulling from listicles, review platforms, and definitional pages, and if you are absent from those specific surfaces, your Google position does not save you.

Why review platforms still decide your fate

Here is the strangest dynamic in the 2026 research landscape. Review platforms have lost most of their human traffic: from 2024 to the end of 2025, G2's organic traffic fell 84.5%, Capterra's fell 89%, and TrustRadius's fell 92%. By click volume, the review-site era looks over.

And yet 88% of review citations in AI Overviews still go to just five review platforms. The humans stopped visiting, but the models never left. Review sites became something new: not a destination, but source material. The buyer never sees your G2 profile. The model does, summarizes it, and hands the buyer a verdict.

This changes the ROI logic of review presence. You are no longer buying traffic from G2. You are buying describability: a structured, third-party-corroborated record of what your product is, who uses it, what it costs, and how it is rated, in exactly the format retrieval systems trust. A thin or outdated review profile does not just cost you a badge. It costs you the sentence the model writes about you.

Why it keeps naming your competitor specifically

Put the mechanics together and the pattern behind "ChatGPT keeps recommending [competitor]" is usually one of four things, and none of them is model bias:

1. They are in the listicles and you are not. A handful of "best [category] tools" roundups feed a large share of citations. If your competitor appears in six of them and you appear in one, the model sees a 6-to-1 vote.
2. They are easier to define. Their homepage says, in one extractable sentence, what they are, who they are for, and what they cost. Yours opens with a metaphor. Models quote sentences, not vibes.
3. Their facts agree with each other. Their name, category label, and pricing read the same on their site, their G2 profile, their directory listings, and press mentions. When sources conflict about you, the model gets less confident and skips you.
4. Their content is fresher. Dated, recently-updated pages beat stale ones, especially in Perplexity. A competitor shipping updated comparison pages monthly compounds this advantage.

Notice that every one of these is fixable without touching your product. That is both the good news and the discipline: this is a content and consistency problem, not a quality contest.

What actually moves the needle

Skip the gimmicks (no, an llms.txt file will not do this for you) and work the four levers the citation data points at.

Get into the listicles that get cited

Find the roundups the engines actually pull from: ask ChatGPT and Perplexity your category question and read the sources they cite. Then work those specific pages: pitch the authors, correct outdated entries, and publish your own honest roundup of the category. Yes, your own. A fair "best tools" page that credits competitors genuinely is heavily mined because it matches the question's exact shape, and honesty is what makes it citable. Our own comparison library exists partly for this reason; see Naoma vs Karumi for the format.

Publish definitional pages

Every important concept in your category deserves a page that answers it in the first two sentences. "What is [category]?" "How does [category] pricing work?" "X vs Y?" These pages are quotable by design. If a model can lift a clean two-sentence definition from your site, you become the source for the entire concept, not just your brand.

Make your entity facts boringly consistent

Pick one name, one category label, one pricing summary, and one one-sentence description. Then enforce them everywhere: homepage, pricing page, review profiles, directories, partner listings, press boilerplate. Corroboration is confidence fuel for retrieval systems. We keep Naoma's core facts identical across surfaces, down to the published pricing, because a model that finds three different prices for you cites none of them.

Keep content fresh and visibly dated

Show updated dates. Actually update the pages: refresh numbers, add the current year, prune dead claims. With Perplexity citing 30-day-fresh content 3.2x more often, a quarterly refresh cycle on your top 10 pages is one of the highest-leverage habits in this whole playbook.

One more thing moves the needle after the citation: what happens when the buyer clicks through. An AI-referred visitor arrives pre-briefed and ready to evaluate, and a static brochure page wastes that intent (our demo conversion benchmarks show how wide the gap gets). The strongest landing experience for a buyer an AI just sent you is the product itself, live, right now; you can see what that looks like if you get an AI demo now.

Run this audit yourself: 10 prompts

You do not need a tool to know where you stand. Open ChatGPT (with web search on) and Perplexity, run these ten prompts for your own category, and log every vendor named and every source cited. Replace the brackets with your real category, ICP, and competitors.

1. "What are the best [category] tools for [ICP]?"
2. "What are the top alternatives to [market leader]?"
3. "[Your product] vs [competitor]: which should I choose?"
4. "Is [your product] legit? What do reviewers say about it?"
5. "How much does [your product] cost?"
6. "What is [category] software and who needs it?"
7. "Best [category] tools for a [company size] company on a budget"
8. "Which [category] vendors support [key capability, e.g. multiple languages]?"
9. "I currently use [adjacent tool]. What should I add for [job your product does]?"
10. "Compare pricing across the top [category] vendors."

Score yourself on three questions. Are you named at all? Is the description accurate, including your pricing? Which sources did the engine cite, and are you present on them? Run the same audit monthly. The engines change fast, and the source list from prompt-by-prompt citations is your literal to-do list: those are the pages to get onto or update.

Expect the results to differ sharply between engines. Remember the 11% overlap: being strong in one and invisible in the other is the norm, not a bug in your audit.

FAQ

Why does ChatGPT recommend my competitor instead of me? Usually for mechanical reasons, not quality ones: your competitor appears in more of the listicles engines cite, is described consistently across review sites and directories, and states what they are in plain, extractable sentences. Audit the sources ChatGPT cites for your category question, and you will typically find your competitor on most of them and yourself on few.

How do I get ChatGPT to recommend my software? Work the retrieval layer first, because it moves fastest: get into the cited category listicles, publish definitional and honest comparison pages, keep your name, category, and pricing identical everywhere they appear, and refresh your key pages regularly. Then maintain your review-platform profiles, since AI engines still lean heavily on them as source material even though their direct traffic has collapsed.

How long does it take to show up in AI recommendations? The retrieval side can shift within weeks: a new page or an updated listicle entry can be cited soon after it is indexed, and Perplexity in particular favors content updated within the last 30 days. The training-data side moves on model release cycles, which is why consistent, widely corroborated descriptions matter; they are what the next snapshot learns.

The takeaway

ChatGPT is not running a popularity contest or holding a grudge. It is doing retrieval over a specific, knowable set of surfaces: listicles, review platforms, definitional pages, fresh dated content, and it names whichever vendor those surfaces describe most clearly and most consistently. Your competitor keeps getting recommended because they are legible there. The response is unglamorous and entirely within your control: run the 10-prompt audit, get onto the sources that are actually cited, say what you are in one clean sentence everywhere, and keep it fresh. Then make sure the click that follows the citation lands on something worth the buyer's intent, because getting named is only half the funnel; what happens on your site next decides whether the recommendation becomes revenue.

Want the buyers ChatGPT sends you to see the product the moment they arrive? Get an AI demo now →