Does list order change what an AI assistant recommends?

Does list position really change AI recommendations?

Yes, in experiments that change only the order. The effect on average is small but real.

The cleanest test is by Baig and colleagues (opens in a new tab). Twelve AI models, from OpenAI, Google, Anthropic and four open models, each chose one of five made-up hotels. The researchers shuffled rating, price, reviews and list position at random, running 3,024 main choice sets per model. Because only the order changed between otherwise similar cards, any position effect is caused by the order itself.

Compared with the first slot, a hotel was recommended 2.1 percentage points less often in slot 2 and 3.7 points less often in slot 5. The authors call list position “a content-free artifact” that “shifts recommendations causally”.

An earlier study by Pfrommer and colleagues (opens in a new tab) at UC Berkeley found the same direction in simulated product search. Even when told to mention the best products first, all the AI models they tested preferred products placed earlier in what they read.

How much is first place worth?

In the hotel study, about $11.7 per night, against $126.4 for a full step up in guest rating.

Converting effects into dollars makes them comparable. Being listed first was worth $11.7 per night in the pooled results, while moving from a 3.9-star to a 4.7-star rating was worth $126.4. A top rating raised the chance of being chosen by 31.6 percentage points, and a high price cut it by 30.0. The other hotel factors are covered in our guide to how AI picks hotels.

So order matters much less than reputation and price. But it is free: the first-listed hotel gained without changing anything about itself. A separate skincare study by Chu and Hou (opens in a new tab) agrees on the scale. Across 14,395 trials with three AI models, position explained 6.5% of the ranking, product details 82.4% and brand name only 1.2%.

Do all AI models care equally about order?

No: sensitivity to order varies a lot by model. An average can hide one model that cares a great deal.

In the hotel study, most models were nearly position-neutral, but gemini-2.0-flash gave the first slot an advantage of about 26 percentage points, roughly ten times the average. A 2026 test by Vishwakarma and colleagues (opens in a new tab) at Sprinklr, a marketing software company, ran 252,000 simulated trials. It pitted two sources against each other in six AI models. Being listed second rather than first was one of four factors that mattered in all six models.

Pfrommer and colleagues also found that the models differed in what drove their rankings: some leaned on product names they already knew, others on order. The research does not let you predict which way a given assistant will lean today.

Where does the order come from in real AI search?

From the search results the assistant retrieves, so search ranking sets the starting order. That is where most of the commercial stakes sit.

The fake-page study by Luo and Chen (opens in a new tab) shows how strong the top slot is. They rewrote real search results so that a page promoted a fake product; the position test used six open models on digital products. Placed first, it fooled the two most vulnerable open models in 27% of cases; placed second to tenth, it fooled them in only 1 to 4%. In their words, “the first page the model reads dominates the recommendation”.

Real data point the same way. In our AI Overview study, the first organic result was cited in 49.5% of AI Overviews, the third in 33.5% and the ninth in 15.5%. Being high in the list an AI reads is linked to being used, though our data cannot show that the position itself causes it. That is one reason classic SEO still matters for AI search.

Will the AI tell you that position influenced its choice?

Almost never. Assistants act on order without saying so.

In the hotel study, list position carried 4.1% of the weight behind the models’ choices. Yet it was mentioned in at most 0.7% of the reasons they gave. Their stated reasons broadly tracked their choices for rating and price, but not for order or review volume.

This matters for monitoring. Reading an assistant’s explanation will not reveal whether your business won or lost on order rather than merit. We look closer at this gap in our guide to trusting AI explanations.

Does a brand’s position in AI answers stay the same?

No: positions in real AI answers move a lot from one run to the next. One snapshot is not enough.

In our consistency study, five runs of the same 20 questions minutes apart, 86.4% of the brands ChatGPT named in every run moved position at least once. ChatGPT’s first pick changed at least once for 80.0% of questions. Position bias may favor whichever brand happens to be retrieved first, and that changes between runs.

What should you do about it?

Track where you appear, not just whether you appear, and work on what sets the order. In practice:

1. Measure your position in AI answers over repeated runs, including how often you are named first.
2. Treat search ranking as part of AI visibility; the pages an assistant reads first carry the most weight.
3. Fix the basics that outweigh order in every study: ratings, clear prices and complete product details.
4. On marketplaces and comparison sites that feed AI assistants, check how listings are ordered and whether you can influence it fairly.
5. Do not trust an assistant’s own explanation as proof that order played no role.

If you want help tracking position across engines, see our generative engine optimization service.

What does the research not tell us yet?

The research proves order effects in controlled tests, but not their size in live AI search. The gaps:

• The cleanest studies use made-up hotels or products and fixed lists; live assistants choose their own sources.
• Effects differ widely by model and by version, and models change often.
• No study we reviewed measures how list position in AI answers affects real bookings or sales.
• Two of the tests come from authors at companies in this market, and several are simulations rather than live audits.
• How much position matters for well-known brands, which models may already know, is not well measured.

Frequently asked questions

Do AI assistants favor the first option they see?

On average, slightly. In a 12-model hotel test, the first slot was worth $11.7 per night, but one model showed an advantage of about 26 percentage points.

Can a business pay to be listed first in AI answers?

The studies here do not cover paid placement. They show that position in the material an assistant reads can shift its choice, which is why that order matters commercially.

Is ranking first on Google enough to be recommended by AI?

No, but it helps. In our study the top Google result was cited in 49.5% of AI Overviews, so more than half the time it was not.

Why does my brand’s position change every time I ask ChatGPT?

Answers vary from run to run. In our test, ChatGPT’s first pick changed at least once for 80.0% of questions across five runs.

Sources

• Baig and colleagues (2026), Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection (opens in a new tab), an algorithm audit of reputation signals in AI-assisted hotel selection, arXiv:2606.16344.
• Pfrommer and colleagues (2024), Ranking Manipulation for Conversational Search Engines (opens in a new tab), arXiv:2406.03589.
• Luo and Chen (2026), One Polluted Page Is Enough: Evaluating Web Content Pollution in LLM Recommenders (opens in a new tab), on web content pollution in AI recommenders, arXiv:2606.13610.
• Chu and Hou (2026), Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems (opens in a new tab), on brand bias and manipulation in AI recommendation systems, arXiv:2606.17443.
• Vishwakarma and colleagues (2026), What Gets Cited: Competitive GEO in AI Answer Engines (opens in a new tab), arXiv:2605.25517.
• Underneath (2026), Do AI Overviews cite the pages that rank?
• Underneath (2026), Ask an AI the same question 5 times: do the brands change?