Vendor Comparison Pages Write the AI Shortlist
I read a vendor comparison page as an input to the shortlist. It is a clipping in the old sense, and our own "us versus them" page is a different document. When a buyer asks an engine who the options are, the answer is assembled from Waves, grids, roundups, and structured reviews that already exist. If we are missing from those pages, we are missing from the list the buyer sees, often before anyone on our team knows the buyer exists.
The account of that mechanism is Vendor Comparison Pages Are Now Writing Your AI Shortlist. Below is how I read one of those pages, and what I try to earn on it.
Who is already building the list
Forrester's The State of Business Buying, 2026, drawn from its Buyers' Journey Survey of nearly 18,000 global business buyers, found that 94% report using AI during the buying process (Forrester). The figure I put in front of a revenue meeting is the later one. Forrester's May 2026 research concluded that generative AI had become the top source of information for enterprise buyers evaluating service providers (Forrester). The shortlist starts as a prompt, and the prompt is answered from comparison content the model can retrieve.
The engine does not call our API to build that list. It retrieves Forrester Waves, G2 grids, analyst evaluations, editorial roundups, and structured comparisons third parties have published. A preprint on agents choosing tools from registries, Advertising to Agents, found that the sales language a provider writes in its listing changes which tool the agent picks. That is about tool registries. I use it as an analogy for buyers' vendor shortlists, and I keep the analogy labeled as one. A missing or vague entry is an empty ad. If the comparison says nothing, the model recommends someone the comparison does describe.
What the description is doing
The current version of that preprint (v3, October 2026) reports a preregistered test on two OpenAI models. Stacking four kinds of praise raised a tool's pick rate by about 43 percentage points, matching or beating a verifiable specification, and with identical listings the first-listed tool was picked about 72 points more often. An exploratory run on five models showed the same direction. A "Sponsored" label dampened the effect, and a system-prompt warning moved one model. The authors call the preregistered results provisional and the models small. An earlier draft (v1) reported larger numbers, among them 17,700 trials and puffery capturing all of the effect. I am not relying on those.
I did not expect praise to match or beat a verifiable specification. In this small study of agent tool choice, it did. The comparison page is doing work. The confident, benefit-framed description is an instruction. Whoever writes that description on the pages the model retrieves has a thumb on the shortlist, including when they did not build the better product. Stuffing superlatives into our homepage is the wrong response. In the study the provider wrote its own listing, which is the opposite of a third-party comparison, and our homepage is a party to the sale. The move is to earn a sentence, on the page the engine already reads, that states the benefit in language we can defend, and to read every existing sentence about us on those pages before we write a new one.
| Source | What the engine can lift | How I weight it |
|---|---|---|
| Forrester Wave or a Gartner Magic Quadrant | Scores, leaders, named criteria | High, because it is independent and structured |
| A third-party comparison | Feature tables, price, pros and cons | High, because it matches extraction |
| G2 or a review aggregator | Sentiment, volume, recency | Medium-high, as corroborated experience |
| Our own site | Claims we wrote about ourselves | Low, as first-party testimony |
| A press release | An announcement without an editor | Minimal, unless a third party picks it up |
I open the Wave or the quadrant if we are in one, then the independent comparisons, then the review grid, and only then our site, to see whether our site agrees. A disagreement is a defect in the entity. The same name, category, and proof on LinkedIn, G2, Crunchbase, the press record, and the structured data is the resolution I try to make boring.
The buying group multiplies the same page
Forrester's buyers' journey data puts 13 internal stakeholders and 9 external participants on the decision. For purchases that include generative AI features, the buying group doubles, 14 members against 7 for purchases without those features. More people means more prompts before the first meeting, all of them able to retrieve the same comparison. If we are absent there, we are absent from 14 conversations that feel, inside the building, like one deal we have not heard about.
Gartner's March 2026 procurement work describes AI across autonomous sourcing, supplier intelligence, and agent-driven source-to-pay (Gartner). The comparison page can feed the procurement system that writes the formal list. I read a page with that use in mind: named criteria, a feature table, a sentence a sourcing agent can lift without our sales deck.
Forrester also finds buyers more likely to engage a provider because of an industry expert than because of an AI tool alone. The tool decides which expert the buyer sees first. AuthorityTech's B2B research synthesis sets Forrester's buyer data beside Muck Rack's finding that 85% of the non-paid share of cited links in its July 2025 edition came from earned media sources (Machine Relations research). That is a share of non-paid cited links. The credibility mechanism is the one PR has used for years. The reader is now a model parsing structured comparison data.
An "Us vs. Competitor" page on our domain, with nothing else, loses that parse. A page on our domain is self-interested. A page in an analyst report or a reported roundup is independent. Earning a place on the pages the engines already trust is the work. Writing our own comparison can still help a human who has already arrived, and it gives the third party a clean source to quote.
The way I organize the work is Machine Relations: earned authority on the comparison surface, an entity that resolves, a claim structured enough to extract, then a measurement of whether the shortlist includes us. I use the glossary for share of citation so the metric stays "how often we are cited against competitors on category questions."
What I do
I list the category prompts our buyers actually use and save the comparison URLs the answers cite. I read each URL for four things: whether we appear, the exact benefit sentence if we do, the criteria the page uses, and any fact that contradicts our site. I correct a contradiction before I chase a new placement. Then I earn coverage in the publications those comparisons already cite, upstream of the roundup.
I measure share of citation on the category questions. Rank and website traffic can stay as operating metrics. They are a different record from the shortlist. If every stakeholder in a 14-person group can receive the same absence, the absence is the pipeline problem.
This is for B2B categories where an engine is asked for a best-of or a comparison before sales is invited. A consumer product with no analyst grid is a different reading problem, closer to reviews and retail feeds. I read whatever comparison the engine retrieves, and I believe that page over our homepage when the two disagree, because the engine will.
A category panel with the prompts disclosed, in which first-party pages outweighed third-party comparisons in vendor selection, and in which praise on those third-party pages lost its effect once specifications were present, would change the spend. I do not have that panel. The registry preprint is a warning about how far a confident sentence can move a model, and it measures neither shortlists nor third-party pages. If specifications start to dominate and our own docs become the retrieved comparison, I invest in the docs and spend less on earning the grid. Until that result exists, I read the third-party page as the draft of our shortlist.
FAQ
How do I decide whether a comparison page matters?
I check whether our buyers' prompts cite it. Forrester's survey puts AI use during buying at 94% of nearly 18,000 business buyers, and the May 2026 research puts generative AI as the top information source for enterprise buyers evaluating service providers. A page those answers retrieve is part of the shortlist. A page they never retrieve is a human asset, and I fund it as a human asset.
Can our own comparison fix our position?
It can help a visitor who is already on our site. On the evidence here, it cannot outvote a Wave, a G2 grid, or a reported roundup. The registry preprint is about provider-written listings, so it tells me little about third-party pages. The engine retrieves independent comparison pages, which Forrester's and Muck Rack's data describe. Our site is first-party testimony. I keep it accurate so the entity resolves, and I earn the independent description separately.
What do I watch in the wording?
Praise that matched a verifiable specification in a small study of agent tool choice. I read the current sentence about us on each retrieved page and ask whether we can stand behind it. A louder sentence on our own domain is a sentence the model may never read. It is reading their page.
How does the size of the buying group change the reading?
Forrester's figures are 13 internal stakeholders and 9 external participants, doubling to 14 against 7 when the purchase includes generative AI features. Each person can run the prompt. Gartner's March 2026 procurement note extends the same pages into sourcing systems. I assume the comparison is read many times, by people and by agents, and I want one consistent description. Fourteen people running the same prompt against a page that omits us is fourteen versions of a deal we do not know exists.
What metric sits next to the page?
Share of citation on the category prompts: how often we are cited relative to the competitors those prompts name. Forrester's point that expert information still closes, while AI decides which expert appears first, is why the metric is citation on the comparison. If the explainer is never retrieved, the traffic metric can sit there looking fine until the shortlist is already written.
About the author
Jaxon Parrott founded AuthorityTech in 2018 and coined Machine Relations in 2024. He writes about earned authority, AI discovery and founder judgment.
Continue reading
The Category Ownership Trap
Owning a category name leaves the buyer's question open. I coined Machine Relations in 2024, and the commercial question is still a separate job.
Is Pay-Per-Placement PR Worth It?
I pay for a placement only when it makes an engine more able to name the company, and I treat the invoice model as secondary.
How to Read LLM Referral Traffic
Read LLM referral traffic as its own channel and as a floor. Seer measured ChatGPT referrals at 15.9 percent against 1.76 percent for Google organic.