Why Do Doubao and DeepSeek Answer Differently?

Set the scene. You hear that buyers now ask AI, so you sit down and try it yourself: open Doubao and ask which manufacturers in Shenzhen make outdoor portable power stations; then open DeepSeek and ask exactly the same words.

The lists overlap only partly. Put the question to ChatGPT in English and a third set of companies appears.

Which brings the obvious question: which one should you optimise for?

The conclusion first: different engines give different answers not because each has its own preferences, but because they were never searching the same library. So the question worth asking is not which engine to court, but which investments every engine rewards. That is where the money belongs.

The boundary too: this article contains no table of "which sites each AI prefers". Most numbers of that kind circulating in the market come from a single crawl or a tool vendor's own account, with inconsistent sampling and different time points. They are not cited here — using them badly is worse than leaving them out. Only two things are covered, both checkable: why the answers differ, and which factors hold across engines.

AI does not "know" — it looks things up

Many owners picture AI as holding a giant list of companies in its head and naming whoever is famous. That is not what happens.

The moment you hit enter, AI does roughly four things: splits the question into several sub-queries, retrieves a batch of pages from an index, extracts usable sentences from those pages, and assembles the answer from them.

Google's documentation calls this query fan-out — one question broken into multiple retrieval actions, each covering a different facet. The mechanism is documented, not industry guesswork.

Hence the key point: every sentence AI produces was just retrieved, not long remembered.

Follow that through and the difference explains itself: two different AI systems sit on different indexes, have crawled different sites, and rank with different models. You are querying the same question in two different libraries. Of course the results do not match.

It is also why so many people ask "I published articles online, so why doesn't AI mention me" — where you published may not be in that engine's retrieval pool at all.

How far apart? This has been measured

This has been measured, and the finding is more extreme than most people expect.

A SIGIR 2026 paper measured the overlap in sources cited by different engines using Jaccard similarity — the value was 0.18. What that means: if two engines each cited 100 sources, only about 18 were used by both.

A second study (arXiv:2603.16138, titled "Answer Bubbles") used 11,000 queries across 4 systems and puts it more directly: different AI systems operate in structurally different information realities.

Scope note: both are academic studies (the first a conference paper, the second a preprint). Their conclusions are bounded by their own query sets and experimental conditions, and they measure overlap in cited sources — not the weight a given engine assigns to a category of site. They cannot be reversed into "engine X only uses these few sites".

Even read only for direction, the number forces a recalculation: if you invest in one engine alone, the effect on your performance elsewhere may be far smaller than you assume.

The answers diverge — but three things hold across engines

The natural reaction so far is to do it five times over, once per engine. But there is a more important finding, and it runs the opposite way.

Although the sources cited barely overlap, research and industry testing of what makes content more citable keep returning the same short list. The three strongest:

• Third-party citations. A study interviewing 100 marketers (NP Digital, June 2026) concludes that third-party citation is the only signal consistent across all six platforms, scored between 4.5 and 4.8. Engines differ in their preferences; every one of them credits other people talking about you.
• Topical relevance plus position in context. A programme of 252,000 controlled trials (six models, 18 factor categories, one variable changed at a time) concludes that what decides whether a passage is cited is how relevant it is to the question and where it sits on the page. Brand names were anonymised throughout, precisely to remove the influence of fame.
• The volume of third-party pages. An industry study (AirOps, "State of AI Search" 2026) reports that about 85% of brand mentions in AI answers come from third-party pages, not from the company's own site.

Scope note: the first and third come from organisations or vendors with a commercial interest and are self-published; they are observational statistics, not independently verified, and support no causal claim. The 252,000-trial study is controlled, but its conclusion is a relative preference under controlled conditions, not a promise of results.

Put the second and third sections together and a counter-intuitive conclusion follows:

What differs between engines is where information comes from. What they share is what makes content worth citing.

You cannot control the first. You can control the second. That is what determines how the budget splits.

So the strategy is two layers, not one engine

On that basis, GEO spending splits into two layers. They are qualitatively different, and running them together is the most common form of waste.

Layer one: the foundation — shared across engines, done once, paying off everywhere

This layer is about what makes content worth citing, which every engine rewards. Four things:

1. Consistent facts. Company name (including the English name), founding date, address, phone, core business — one version across the web. Inconsistency is itself a penalty — when AI cross-checks, it discounts both versions.
2. Actually answer questions. List the 20 questions your buyers ask most and check, one by one, whether your site answers them head-on. This does more than 100 more generic industry articles.
3. Make content extractable. One point per paragraph, under subheadings; define a concept before using it. Note that structure means how information is organised, not font size and colour — the two get conflated constantly.
4. Exist on third-party ground. Industry media, B2B platforms, communities, video, encyclopaedias — the places others mention you are the source of that 85% of brand mentions.

Do these four and you gain a chance of being found whichever engine a buyer turns to — Doubao, DeepSeek or ChatGPT. And they do not stop working when one engine changes its algorithm.

Layer two: placement — cover the divergence by source type, not by guessing at engines

One counter-intuitive rule to state plainly: place by source type, not by engine.

The reason is simple: you cannot guess accurately. Each engine's source mix shifts with version, region and time, and any list of "engine X favours site Y" can be stale within months. Spending against such a list means betting the budget on a rule that moves.

Source types, by contrast, are stable: own website, encyclopaedia entries, industry media, content communities, short video, storefronts on B2B platforms, developer and technical communities — these pools have barely changed in a decade. Engines change; the pools do not disappear.

So the correct move for layer two is: find where your buyers go to learn about suppliers, and be present there. The more pools you cover, the higher the chance any engine retrieves you. It also explains why a single channel loses over time — not because the channel is useless, but because it covers one pool.

Three common questions

We only sell domestically. Do we still need to bother with ChatGPT?

Look at where your buyers are, not at how famous the engine is. If your customers are all domestic and decide in Chinese throughout, the focus does belong on Chinese engines. One case to watch: for exporters or companies preparing to go overseas, the overseas buyer has probably already asked in English before sending a formal inquiry. Missing English content costs you not a ranking, but exposure across the entire research phase. The test is simple: ask three existing customers what they looked up before contacting you.

If every engine answers differently, how do I measure my "ranking"?

Do not measure ranking. Measure appearance rate. An AI answer is not a list; there is no stable first and second place, and asking the same question twice can produce different results. The workable method: fix a list of 10 to 20 questions your buyers genuinely ask, put them to each engine every two weeks, and record three things — whether you appeared, whether what was said about you was accurate, and which competitors appeared. Three months of that gives you a trend worth more than any screenshot of a ranking.

Should we start with whichever engine has the most users?

Total user count and whether those users are your buyers are two different things. B2B purchasing behaviour differs sharply from consumer behaviour: the questions are longer, more specific, and more often verified in professional channels. Rather than total users, ask how many of your own customers actually use that tool — a number you can estimate by asking five of them, and more accurate than any industry report.

Where this leaves you

To pull this together: "which AI should we optimise for" may be the wrong question to begin with.

What differs between engines comes from sources; what they share comes from whether the content is worth citing at all. You can neither predict nor control the first. You can control the second, and doing it once pays off everywhere.

So the practical order is foundation first, pools second. Rushing to study what one engine likes before the foundation is set is painting the walls before building the house.

Start with one small thing: take the three questions your buyers ask most, put each to Doubao and to DeepSeek, and lay the answers side by side. You will see two things at once — which companies appear consistently, and which only show up in one engine. That single comparison reveals what you are missing better than any proposal.

Yihang Network (Shenzhen) previously worked in cross-border B2B and Alibaba.com-related services, and has been extending into GEO (Generative Engine Optimization) since 2024. Our diagnostics follow the same order — foundation, then placement: first whether brand information is consistent and whether the content actually answers buyer questions, then how coverage looks engine by engine.

Data and effect boundaries: the citation overlap figure (SIGIR 2026 paper, Jaccard 0.18) and the Answer Bubbles study (arXiv:2603.16138, a preprint, not peer-reviewed) are academic research, with conclusions bounded by their respective query sets and experimental conditions. The NP Digital six-platform signal comparison (June 2026, 100 marketers) and the third-party mention share in AirOps "State of AI Search" 2026 come from self-published research by organisations with a commercial interest; they are observational, not independently verified, and support no causal claim. The 252,000-trial study (SIGIR 2026 accepted paper "What Gets Cited") reflects relative preference for candidate content under controlled conditions, not real-market performance. None of the above represents the results of any specific company. The methods described raise the probability of being retrieved and cited; they do not guarantee that a company will appear in any AI answer, and they promise no timeline or business result.

Sources

1. SIGIR 2026 paper: overlap in cited sources across generative search engines (Jaccard 0.18)
2. arXiv:2603.16138, "Answer Bubbles", 11,000 queries across 4 systems (preprint, not peer-reviewed)
3. NP Digital: assessment of AI search signals by marketers, June 2026, 100 interviews
4. AirOps: "State of AI Search" 2026 (share of brand mentions on third-party pages; share of citations from community platforms)
5. SIGIR 2026 (accepted): "What Gets Cited", 252,000 controlled trials, 6 models × 18 factor categories
6. Google Search Central: "AI features and your website" (documentation of query fan-out)