When a Buyer Asks AI How You Compare

A salesperson at a packaging machinery maker told me this. A buyer asked Doubao which of two companies was better, and AI gave three reasons, two of them "YY has shorter lead times" and "YY holds CE certification". He said their lead time was shorter, they had CE too, and they got it first. AI did not mention them once.

His first reaction was that AI played favourites. The actual reason is more blunt: AI is not comparing who is better. It is comparing whose content contains comparable fields.

The competitor's page read "standard models, 25-day lead time, CE certificate number XXX, 24-month warranty". His page read "excellent quality, attentive service, flexible lead times". The first goes into a comparison table; the second does not. So this was not a loss on points. He was never entered. What does that have to do with GEO? A great deal — comparison questions are the last question in B2B purchasing, and whoever is absent from that question has wasted every earlier bit of exposure.

The boundary first. This article is about whether AI holds your data at the moment a buyer compares you with a competitor, and it does not apply to companies that have not reached the comparing stage — that is an exposure problem, and being seen has to be solved first. Nor does every category run on comparison: for highly custom equipment, buyers often compare only one supplier, because nobody else can make it. Finally, everything below only raises the probability of being included in a comparison. It cannot guarantee an answer favourable to you. Generative engine optimization has no means of steering the outcome, and anyone who guarantees you will beat a rival has ended the conversation.

Comparison is the last question in the purchase

Purchasing is not a single decision. It is a string of questions. First "what is this thing called" (category), then "who can make it" (supplier), then "will this specification work" (specification), and last "which of these do I pick" (comparison).

The earlier questions all have alternative routes: the buyer can ask a peer, walk a trade fair, call a trader. Only the last one is increasingly taken straight to AI, because it is fast and looks neutral.

The answer to that question has a shape: it is a list, ordered, with reasons attached. Not "a certain company is decent", but "A has the price advantage, B is steadier on lead time". That shape dictates that your content has to fit into a table, or you do not even occupy a slot in the answer.

One aside: different AI systems often reach different conclusions on the same comparison. The SIGIR 2026 paper measuring citation overlap between systems puts the Jaccard coefficient at only 0.18 — for one comparison question, the pages cited by Doubao, DeepSeek and ERNIE Bot barely intersect, so the conclusions can naturally conflict. A buyer who asks two systems gets two versions. This is why GEO has to test several engines rather than one.

AI compares on fields, not adjectives

Take an AI comparison apart and the structure is fixed: several candidates, several dimensions, a value for each dimension. See that structure clearly and the focus of AI search optimization changes.

Where do the dimensions come from? From the pages that get retrieved. If both pages state lead time, certification, warranty and minimum order quantity, the comparison runs on those four. If only one page states them, that page supplies the dimensions unilaterally — the other party is blank on all four, and blank in a comparison usually reads as a disadvantage.

That is where the three reasons came from. Not favouritism; one side's data was all it had.

So at this layer the task in AI search optimization changes: not writing better adjectives, but filling in the numbers and facts that can be compared side by side. In B2B, what actually gets compared is roughly this:

• Standard lead time and rush-order lead time, in days
• Minimum order quantity and the volume tiers
• Key tolerances, materials, process limits
• Certification name, number, scope and validity
• Warranty period and how after-sales response works
• Pricing basis: tax included or not, freight included or not, how long the quote holds

There is no column in that table called "excellent quality". Not because quality does not matter, but because it cannot be filled into a cell. The first thing generative engine optimization usually has to change is exactly those adjectives.

Without comparable fields, AI skips you

Being skipped is quieter than being wrong. Write something wrong and you get feedback; be skipped and you learn nothing. A large share of "we did the work and nothing happened" cases in GEO stall at exactly this step.

One study shows what a place in the comparison is worth. The KDD 2024 paper ran an experiment: adding citations to a result originally ranked first dropped its selection rate by 30.3%, while adding citations to a result originally ranked fifth raised its selection rate by 115.1%. Read the other way it holds too — in a comparison setting, the quality of the material can rewrite the ranking rather than merely reinforce it.

That conclusion matters because ranking is not the only variable. Ranked low originally, with comparable and evidenced material you can still be lifted; ranked high with empty material, you get pushed down. This is one of the few points in AI search optimization that experimental data can spell out.

The University of Toronto preprint measured AI's preference for third-party sources at 53% to 95%, depending on query type. Comparison questions are precisely where that preference is strongest, because the buyer is asking what other people think of these companies.

Which raises the next question: is writing it yourself enough. Many people stall in GEO because they assume the website is the end of it.

Third parties decide whether you reach the table

It is not. Fields you write yourself only solve "the data exists". Whether it is accepted is another layer. In AI search optimization, changing only your own website never reaches that layer.

On that layer, AirOps' "State of AI Search" 2026 gives an observation: when a brand is mentioned, 85% of the sources are third-party pages rather than the company's own site. A Semrush study of more than 540,000 AI answers supplies another: 62% of AI citations never produce a brand mention. Fasten the two together and the picture is this — AI reads third-party content heavily, and often finishes reading without naming anyone.

In a comparison setting that means: when a buyer asks which of A and B is better, AI is more likely to cite a trade-press comparison, a buying guide, or real usage feedback in a forum than the "why choose us" page on your own site.

So the third-party layer is not extra credit. It is the ticket in. It has three main parts:

One, verifiable customer cases. Not "the customer was satisfied", but the industry, the operating conditions, the problem solved, what was delivered. Customer names can be anonymised; the facts cannot be soft.

Two, presence on the industry side. Association directories, exhibitor records, trade-press coverage, participation in drafting a standard or white paper. The characteristic of this material is that someone else wrote it, which puts its credibility a grade higher.

Three, reviews and Q and A. Thinner in B2B than in consumer goods, but present. In G2's 2026 survey, 45% of B2B buyers said citing review sites was the most confidence-building signal. That is an overseas figure; the equivalent in China is real reviews in industry communities and vertical platforms — different carriers, same logic.

B2B buying is already looking for a second source

One premise worth stating: when a buyer asks how you compare with a competitor, it is often not to eliminate anyone but to find an alternative. That decides how the GEO objective should be set.

A peer-reviewed paper in Computers and Industrial Engineering, August 2026, concludes that dual sourcing delivers steadier profit and better risk hedging across levels of disruption risk. In McKinsey's "Supply Chain Risk Pulse 2025", 39% of respondents practise dual sourcing, 45% are raising safety stock and 33% are moving toward near-shoring. The industry also cites a concentration rule of thumb: no single supplier above 70%.

Those figures say one thing: a buyer will almost certainly look at several companies at once. They are not seeking a perfect supplier. They are assembling a supplier portfolio.

That is good news for the party being compared. A common industry observation is that a second source typically has to quote 10% to 20% higher to cover switching cost — meaning buyers will pay a premium for having a reliable alternative.

For anyone doing GEO the conclusion is: the objective is not to be the only company recommended, which is neither realistic nor necessary. The objective is to appear in that candidate table consistently, with a value in every dimension. In export-heavy places such as Shenzhen and Dongguan, portfolio buying is more common than single sourcing, which makes this hold even more firmly.

Four things you can do

In concrete terms, GEO at this layer comes down to four actions.

First, build your own comparison table. Lay out those six dimensions and fill in your real values. A cell you cannot fill is information you have to go and get, not a place to hide. The table doubles as an internal management tool — many companies get halfway and discover their own lead-time definitions disagree between departments.

Second, put the fields in the body text, not in an image. Parameters inside a JPG cannot be read. Make them a table or a list, written as text.

Third, get third parties to use the same fields. Carry lead time, certification and acceptance terms into customer cases; carry specific capacity and process detail into trade coverage. Where third-party material and your own site disagree, both are discounted.

Fourth, accept being compared. Some companies will not publish lead times or minimum order quantities for fear competitors will see them. The concern is understandable, and the price is that you remain the blank cell in every AI comparison. A compromise is to give a range rather than a precise value — "20–30 days for standard models" — which gets you into the table without showing every card.

Three common questions

My price is higher than competitors. Would publishing it not hurt me?

Leave price out and AI will cite somebody else's figure on that dimension, which leaves you worse off. What works is stating the price basis and its scope rather than an absolute number — "tax and freight included, MOQ 500 pieces, quote valid 30 days" is itself competitive information, and it explains why you cost more. The role of GEO here is to make "expensive for a reason" something that can be read.

Will buyers use AI's comparison to push our price down?

Yes, and it is already happening. The response is not to make AI say less but to make your own basis clear enough that a buyer can see where the differences lie. A vague quote is the easiest thing to compare.

A competitor is already on AI's list. Is there still a chance?

Yes, but it depends which kind of list. A comparison answer is assembled dynamically and may draw on different sources on each ask; a fixed directory is hard to change because it is itself a page. So establish which one you are facing first — the same logic as the three-state diagnosis described earlier.

Where this leaves you

Back to the packaging machinery salesperson. What we did afterwards was simple: put his lead time, certificate numbers, warranty period and acceptance standards into a table on the product page, then had two long-standing customers say the same things in their own channels. Three months later the buyer asked the same question again and AI mentioned them, on the grounds of "20-day lead time, CE certified".

Note that both facts AI cited were things he already had. What changed is that they moved from "only the salesperson knows" to "written on the page, and others say the same".

Yihang Network has worked on GEO since 2024, having previously been in cross-border B2B and Alibaba.com-related services. The logic of scoring B2B purchases by dimension came from there — buyers on Alibaba.com were already comparing parameter tables, and nothing changed when the venue moved to AI. So when we look at a client we ask one question first: if your buyer took you and two competitors to AI tomorrow, which of your numbers would AI be holding?

Data and effect boundaries: the dual-sourcing conclusion from Computers and Industrial Engineering, 19 August 2026 (DOI 10.1016/j.cie.2026.112192) is a peer-reviewed paper whose scope is supply-chain risk management rather than purchasing decisions in general. The KDD 2024 figures of 30.3% and 115.1% are relative changes under controlled experimental conditions, not a guarantee of online mechanism. The University of Toronto range of 53–95% and the SIGIR 2026 Jaccard of 0.18 are preprint or paper findings bounded by the query set tested. AirOps' 85%, Semrush's 62% (from more than 540,000 AI answers) and G2's 45% are vendor or platform-published observational statistics, not independently verified, and describe differences between groups rather than causation. McKinsey's 39%, 45% and 33% are survey figures at n=100, relayed by third parties; the 70% concentration rule and the 10–20% second-source premium are industry rules of thumb, not authoritative statistics. All of the above are mechanism descriptions or observational statistics and do not represent any company's actual result. The methods described raise the probability of being retrieved, cited and included in a comparison; they guarantee no favourable conclusion from AI, and promise no timeline and no business outcome.

Sources

1. KDD 2024: DOI 10.1145/3637528.3671900 (adding citations lowers a result ranked first by 30.3% and raises one ranked fifth by 115.1%; peer-reviewed)
2. Computers and Industrial Engineering, 19 August 2026, DOI 10.1016/j.cie.2026.112192 (dual sourcing delivers steadier profit and better hedging across disruption risk levels; peer-reviewed)
3. MIT Sloan Management Review (two problems with multi-sourcing: alternatives often sit in the same region as the main supplier, and managing two suppliers costs more)
4. McKinsey: "Supply Chain Risk Pulse 2025" (39% dual sourcing, 45% raising stock, 33% near-shoring; n=100, relayed by third parties)
5. G2: 2026 B2B buyer survey (45% find citing review sites the most confidence-building; platform-published)
6. AirOps: "State of AI Search" 2026 (85% of brand mentions from third-party pages; vendor-published, observational)
7. Semrush: "ghost citations" study across more than 540,000 AI answers (62% of AI citations never produce a brand mention; vendor-published)
8. University of Toronto: arXiv:2509.08919 (AI answers prefer third-party sources 53–95%; preprint, not peer-reviewed)
9. SIGIR '26: study of citation overlap across systems (Jaccard 0.18; peer-reviewed)