DeepSmith

Nov 26 · AEO & AI Visibility

19 min read

Does Position Matter in AI Answers? Rethinking Rank for AI Visibility

Avinash Saurabh
Avinash Saurabh · CO-Founder & CEO
Layered monochrome answer cards with a slightly shifted top card, node lines and small bar fragments behind the cover line about position in AI answers.

Being listed first in an AI answer might help, but the evidence is not strong enough to make it your main goal. We grade the evidence on AI answer position as mixed. One small 2026 study found that people picked the first item in Google's AI Mode most of the time. Several repeat-answer studies found that the brands named in an answer, and the order they come in, change from one run to the next. So first place looks like a useful thing to notice and a poor thing to build a whole strategy around.

If you've been asked "where do we rank in ChatGPT?", this piece gives you a way to answer that honestly. We'll look at what "position" can mean, what the best available studies show, what the official documentation says, and what to track instead of a single rank number.

What people mean by "position" in an AI answer

Before you can say whether position matters, you need to say which position you mean, because the phrase gets used for at least five different things.

The first is brand order inside the answer. That's the brand named first in a recommendation list, and it's the meaning this article is mostly about. The second is where the AI answer sits on the screen, for example near the top of a search results page. That's about layout and says nothing about which brand comes first inside the answer. The third is citation order, meaning how prominent the links attached to an answer are. A citation is a linked source, not a recommendation, so a high citation doesn't mean a high brand rank.

The fourth is how much of the answer text is tied to a given source and where that text falls. Researchers sometimes weight this, but those weights are design choices and not measured human attention. The fifth is input order, which is the order in which a question names the brands being compared. That's an effect of how someone worded the prompt, and it isn't the same as which brand the answer shows first.

When people say "ranking in ChatGPT answers" or "AI search rank," they usually mean the first one, a brand's order in a list. It helps to know that none of these phrases names a documented, universal rank metric. OpenAI's help documentation says ChatGPT ranks search results using multiple factors and that placement isn't guaranteed. That statement is about search results, and it doesn't define a stable number for where a brand lands in every generated answer.

The evidence that position can matter to people

The strongest case for caring about position comes from a 2026 behavioral report from Citation Labs, produced with Eric Van Buskirk and the Clickstream Solutions team. It was a remote, unmoderated think-aloud study. Forty-eight U.S.-based participants worked through major-purchase tasks, and after data cleaning the report counted 185 tasks.

In the rank-choice table for AI Mode, which covers 137 observations, 74% of people chose the item ranked first as their top pick. The report gives a mean rank of 1.54 for first choice, and 10% chose something ranked third or lower. About 26% overrode the order, usually because they already knew and trusted a brand. Even among those people, 81% still ended with a shortlist that came from the AI's output.

The same report points to something other than order. It says the AI's product description drove the decision in 48% of AI Mode tasks, compared with 6% in standard Search sessions. It also says 74% of AI Mode shortlists came straight from the AI output, and 64% of participants clicked nothing during the task. Most buyers skimmed the answer. So in this study, people leaned on the first item and also on how the answer described the products.

That's useful, but it comes with limits you should keep in mind when you repeat it to anyone else.

  • It's a small, task-based study, not a controlled test. It doesn't show that moving the same item from second to first changes what people choose.
  • The available report doesn't establish that the order of the AI's list was randomized, so a strong brand may simply have been listed first and chosen because it was strong.
  • The table for this result uses 137 observations, while the report's no-click table uses 149 and the full cleaned set is 185 tasks. Those denominators differ, so be careful about mixing them.

The honest reading is that 74% is an association in one study, in one product, on one kind of task. It isn't a click-through rate, it isn't a general user preference, and it isn't proof that first position causes the choice.

The evidence that order and inclusion are unstable

The second body of evidence asks a different question: if you ask the same thing again, do you get the same brands in the same order? The answer from the repeat-answer studies is that often you don't.

The Asked Thrice repeatability study

Asked Thrice published a study on August 23, 2026, based on 4,500 model API calls made that day. It covered 30 categories, used buyer-question templates across three providers, and repeated each question five times. The calls used budget-tier models, with no system prompt, no tools and no web search.

It reported three numbers that are easy to mix up. In 15.0% of question-and-provider cells, all five runs named exactly the same set of brands. In 9.6% of cells, all five runs gave the same brands in the same order. And its "top-one stable" measure was 75.1%, which means the first-named brand was the most common first-named brand across the five runs in that cell.

Those are three different questions. The 75.1% figure is not the rate at which the first brand repeats unchanged, and the 9.6% figure is about a whole list in identical order, not just the first item. The study also reported a mean Jaccard overlap of 0.54 for brand sets across 900 cells, and 49.8% of brands showed up in only one of the five runs.

The study has limits that matter. It used templates and not real user prompts, it measured a single day, and it normalized brand names with a heuristic. Its list extractor failed to find a list in 386 of the 4,500 answers. OpenAI and Gemini were run at temperature zero while Claude used the API default of 1.0, so comparing providers is muddied by the settings. It also says nothing about people in the live consumer apps.

The EWR Digital variance study

EWR Digital's version 2 AI Answer Variance Study, dated November 5, 2026, looked at 28,752 answers from 256 prompts across five engines: ChatGPT, Perplexity, Gemini, Claude and DeepSeek. The answers were collected between January 13 and October 15, 2026. In that sample the focal brand was named in 26.9% of answers and its own domain was cited in 21.6%.

Run-to-run agreement was 0.840 for being named and 0.846 for being cited. Those figures describe whether a brand was present or cited across repeats. They don't say which one matters more to a buyer, and they don't say anything about order.

Treat this study with some care. It's observational, it used one tracking instrument and seven non-random commercial sites, and the engine versions changed while the data was collected. The version 2 report also corrected an earlier claim, and it withdrew the idea that being named was 2.7 times more stable than being cited, because that comparison divided measures on different scales. If you see that 2.7 figure quoted anywhere, it has been withdrawn. And its "identical cited-source set" measure leaves out any pair of answers where either one cited nothing, so the denominator matters.

A peer-reviewed look at inclusion

A study in the Journal of Marketing Analytics, published October 31, 2026, looked at 50 brands and 250 queries across five industries and three language-model systems. Queries were repeated five times in February and again in September 2026, giving 3,614 and 3,750 scored answers. It found a very strong correlation between the two dates in how often brands were included (Spearman 0.994). Agreement across the different systems was 57.2% in February and 60.8% in September.

That sounds like good news for stability, and for inclusion at the level of a whole brand set, it is. The same study warns that appearing in an AI-generated supplier list doesn't show that buyers considered or chose a brand, and it didn't measure consumer outcomes. Its author is affiliated with Rankfor.AI, a company that builds AI brand-intelligence tools, which is worth knowing when you weigh it.

Put together, these studies suggest something fairly consistent. Over many runs, whether a brand tends to be included can be fairly steady, while the exact order of a single answer is much shakier. If you want to understand why that happens, we explain the mechanics in why AI answers change from run to run.

Input order is a different effect

One more study is easy to misread, so it gets its own section. An Opascope analysis dated October 23, 2026, asked a frontier model with live web search to compare two named real brands, and it ran each comparison in both input orders with repeated sampling.

The focus brands won 62.8% of comparisons when they were named first in the question and 53.4% when they were named second. When the model showed no preference, the first-named brand still won 63.3% of the time. The reported odds ratio was 2.69, based on vote pairs of 344 and 128, with a 95% interval of 2.19 to 3.32. The page says it didn't measure how buyers order brand names.

This is about the order of names in the question you ask, and the result is a model recommending one of them. It isn't evidence that a person is more likely to choose the brand shown first in an answer. What you can take from it is a testing habit. If you check recommendation prompts that name two brands, keep the wording and the brand order fixed from run to run, or test both orders on purpose, so you aren't measuring your own phrasing.

What the official documentation says

The platforms themselves don't describe a brand rank, and it's worth reading what they do say.

OpenAI's ChatGPT search help page says web-search responses may include citations and that a Sources panel, when available, lists cited sources and other relevant links. It warns that results and citations may be incomplete, outdated or incorrect. It says search results are ranked using multiple factors but that placement isn't guaranteed. It also says ChatGPT may rewrite queries using saved memory when that's turned on, and may use approximate IP location or an optional device location for local results. Those inputs alone can make two people see different answers, which is the subject of our piece on personalization and location in AI answers.

Google's AI features documentation, last updated December 10, 2025, says AI Overviews and AI Mode may use query fan-out, which means issuing several related searches across subtopics and data sources. It says the systems may use different models and techniques, so the responses and links shown can vary. A page needs to be indexed and eligible to appear in Google Search with a snippet to be eligible as a supporting link, and Google lists no extra technical requirements. The page doesn't define a first-brand rank.

OpenAI's API web search documentation draws one more distinction. Inline citations show the most relevant references, while the sources field returns the complete list of URLs the model consulted, which is often larger than the citations. That's API documentation and not a description of every consumer answer. Still, the practical point holds: the citations you can see and the sources the model read are not the same list, and neither one is a brand rank.

Research metrics that weight position are proxies

You may run into research that does weight position. The 2024 GEO paper by Pranjal Aggarwal and coauthors introduced a metric called Position-Adjusted Word Count. It measures how visible a source is by weighting the text tied to a citation according to where that supporting sentence appears in the generated response, with earlier sentences counting for more through an exponentially decaying factor.

The paper's reasoning is that earlier sentences are more likely to be read, but the weighting is a design choice and not a human attention curve that the paper measured. Its other measure, called Subjective Impression, relies on model-based evaluations and not a human study of brand order. The paper tests ways to improve source visibility, and it doesn't show that the first brand in an answer gets more clicks, better recall, more trust or more purchases.

So it's fine to use metrics like this as working proxies, as long as you call them proxies. Citation position isn't the same as brand-list position, and a model's visibility score isn't the same as what a buyer did.

A fair verdict on whether position matters

Putting it together, here's where we land, and we'd call the grade mixed.

AI answer position may matter to people. One study found a strong link between first place and selection in AI Mode, and a lot of buyers skim and take the shortlist they're given. At the same time, nothing reviewed here isolates the effect of position by itself, and the research on repeated answers shows that the exact order of a list is often not stable. Most of the evidence on order comes from API calls made without web search or from a model's input order, which are different from what a person sees in a live app.

The takeaway isn't that rank doesn't matter. It's that a single rank number doesn't describe something solid enough to report as your AI search rank, and chasing it can lead you to read noise as progress. That's especially true if the number comes from one screenshot, one run, or one prompt.

What to track instead of one rank number

A better approach is to track a few signals over repeated runs, with each one defined clearly and with a stated denominator. The table below shows what each signal tells you and what it can't prove.

SignalWhat it tells youWhat it does not prove
Mention rateHow often the brand is named across a defined set of prompts, runs, platforms and dates.That the mention is favorable, prominent, remembered or chosen.
Framing and sentimentWhether the answer describes the brand accurately and positively, neutrally or negatively, and what role it gives the brand.That people trust or act on the description. The Citation Labs study suggests framing can matter in its task setting, but it doesn't show a universal effect.
Citation rateHow often the brand's pages are linked as sources under a defined collection method.That the brand is named, recommended, read or chosen.
Position distributionHow often the brand is first, second or later when it appears, across repeat answers.A stable, platform-wide AI rank, or the effect of moving the brand to the top.
Top-one frequencyHow often the brand appears first across repeated runs of one fixed prompt.That the whole list is stable, that users respond to it, or that it affects revenue.

Each of these is a visibility measure, not proof of business impact. Still, inclusion, accurate framing and citation presence are easier to interpret from month to month than one ordinal rank, and you can keep a record of all of them without special tools. The difference between a mention and a citation is worth getting right, and we cover it in how to tell if AI is citing or only mentioning you and in AI citations vs brand mentions.

How to run this with your own prompts

Here's a simple way to start, and it works in a spreadsheet.

  1. Pick ten or so prompts your buyers actually ask, and keep the wording the same every time.
  2. Run each prompt several times on each platform you care about, on the same day, and write down every brand named, in order.
  3. Record the platform, the prompt, the date and the number of runs, so every figure has a denominator.
  4. For each brand, work out how often it appears, how often it appears first, and how often your pages are cited.
  5. Read the framing of at least a few answers and note whether the description is accurate and how favorable it is.
  6. Repeat on a regular schedule and compare trends, not single answers.

If your prompts compare two named brands, run both orders, since the input-order result above shows that your wording can tilt the outcome. Tracking mention rate and citation rate over time is also a more honest way to report progress to a boss who asked about rank, because you can say how often you appear and how you're described, with the sample size attached.

DeepSmith's AEO module does this kind of repeated checking for you. You define the prompts, it checks them on a schedule, and it reports mention rate, citation rate, share of voice and sentiment per prompt, with the full answer history kept so you can read how the brand is described. It tracks the engines listed on your plan, and the Enterprise plan covers all ten.

How to talk about this with your team

Sooner or later someone will ask for a rank. You can give them something honest without brushing the question off.

Start by agreeing on which position they mean. If it's order within the answer, say that you can report how often the brand is named first across repeated runs, and that you'd call that a distribution and not a rank. If they mean a screenshot from last Tuesday, explain that a single answer is one draw from a range, and that the studies above found that full lists in identical order were rare in at least one test.

Then offer the signals that hold up better. Say how often you're mentioned, how the answer describes you, and how often your pages are cited, and give the number of prompts and runs behind each. When the rate moves, you'll know whether it's a real change or ordinary variation, and the pieces on handling answer noise and stabilizing citation volatility can help you tell the difference.

You can also say plainly that being first may still be nice to have. There's no reason to ignore it. The point is to keep it as one secondary observation next to the others and not the number everything is judged on. If you want the wider picture of how AI search works and why these measures exist, our guide to what AI visibility is is a good place to start, and the best AI visibility tools roundup covers the tools that measure it.

What would change this verdict

The grade would move up if independent studies changed only the displayed order of otherwise identical AI recommendations, at random, and then measured attention, recall, clicks and choices. It would also help to see larger and more varied groups of participants, and results that held across platforms and across different kinds of tasks.

A good study would also keep the different kinds of position apart: order inside the answer, placement on the screen, citation prominence and input order. Until that exists, the first-position finding is worth reporting, but it shouldn't be treated as a general rule that applies everywhere.

The engines change too. Google says its systems may use different models and techniques, and the studies above collected data while engine versions moved, so even a solid result from today could look different in a year. That's another reason to track over time and not to lean on one reading.

If you'd like to see how your own brand shows up across repeated runs, you can start a free DeepSmith trial and set up your first prompts. The trial is 7 days, and it gives you real data to look at before you decide anything.

Frequently asked questions

Does being listed first in an AI answer matter?

It may. Citation Labs' 2026 task study found that 74% of AI Mode rank-choice observations selected the first-ranked item. That study doesn't establish a general causal effect, and it doesn't show the same thing happens on other platforms.

Is there a stable AI search rank for a brand?

The sources reviewed here don't establish one. Repeat-answer studies find that brand sets and their order can vary, and official documentation describes changing retrieval and response behavior and not a guaranteed placement for a brand.

Are citations and recommendations the same signal?

No. A citation is a linked source, while a recommendation or mention is content in the answer text. OpenAI's API documentation also separates the inline citations that are displayed from the broader list of sources the model consulted.

What should a marketer track if not just rank?

Track how often your brand is mentioned, how the answer describes it, how often your pages are cited, and how positions are spread across prompts and runs. Record the platform, prompt, date and number of runs, and treat all of it as visibility data and not as proof of buyer behavior or revenue.