November 2026

AI Search Impressions: The First Standard Metric for AI Search

How brands can count AI search exposure the way they count every other channel.

Introduction

How do we know we’re getting results?

1 minute read

“Rather than replacing one black box with another, we must translate vanity metrics into defensible, incremental business impact.”

Hitesh WadhwaniCo-founder of Google Meridian

AI search is rewriting the buyer’s journey

The buyer’s journey is experiencing the biggest shift since the creation of search engines. Buyers now turn to ChatGPT, Claude, Gemini and Grok, among other AI platforms and personal agents, to discover, compare, and choose brands, guided by personalized recommendations that arrive in seconds.

Marketing leaders have noticed and are spending accordingly. 86.6% say AI search is important or a top priority, yet only 23% report high confidence in how they measure it (AirOps AI Search Reality Index). Spending has moved faster than measurement.

The difference between AI search spending and attribution leaves marketing leaders with an uncomfortable choice. They can invest heavily without data that shows what it returns, or invest cautiously and cede ground to competitors who move first. For most marketers, growth targets are rising and budgets are staying flat, so neither option holds up for long (AirOps AI Search Reality Index).

What's missing is the measurement that every other channel already has: a shared, standardized count of how often buyers see a brand. Television has ratings, display has the viewable impression and search has the click. AI search has nothing comparable yet.

That's why the AI Search Measurement Standard was created. It brings together data providers, measurement partners, and brands to establish a shared standard for how AI search is measured.

AirOps started it with three partners: Paramark, a marketing measurement company; Chime, the consumer fintech brand; and Hitesh Wadhwani, creator of Google Meridian, Google's open-source marketing mix model. Since the announcement, more than a dozen brands have signed on as supporting participants, along with voices in AI search like Kevin Indig, author of Growth Memo; Cyrus Shepard, an SEO and AI search consultant; and Ross Simmonds, CEO of Foundation.

The group's first output is the AI Search Impression, an estimate of how often buyers see a brand in AI answers each week. This report explains why the AI Search Measurement Standard is starting with AI search Impressions, how the methodology works and where the effort stands now.

Get involved

Join the AI Search Measurement Standard and help shape it. Leave your work email and we will be in touch.

Chapter 1

Almost every major channel started out unmeasurable

7 minute read

AI search is not the first channel to ask for budget before it could prove its worth. Nearly every major marketing channel has followed the same pattern of evolution: the audience arrived first and measurement caught up later.

It's how most new channels begin. They become dependable investments once a shared unit exists, and an industry group usually has to push for it.

From first ads to a shared measurement unit

Tap a row to see its dates

  1. the likelihood-to-see impression
  2. Ratings
  3. Ad impression
  4. Ad impressions and clicks
  5. the download
  6. the standardized ad impression (adopted)
  7. the attributed sale
  8. None yet

Out-of-home (OOH)

Created in 1830sMeasurable in 1933

Billboards date back to the 1830s (OAAA), yet for most of their history advertisers bought them on rough estimates of how many people might pass by. In 1933, three advertising trade groups formed the Traffic Audit Bureau, now Geopath, to audit those traffic counts. It was a first step: the audits showed how many people could have seen a billboard, not how many likely did.

The unit the industry still uses came decades later. Around 2010, the bureau replaced raw traffic counts with the likelihood-to-see impression, which also carries audience demographics (Adweek). Independent oversight followed. The Media Rating Council, a nonprofit that accredits media measurement, finalized US out-of-home impression standards in April 2024 and added audience requirements in December 2025 (Media Rating Council).

Television

Created in 1941Measurable in 1950

Television sold ads before anyone could count who was watching. The first legal US TV commercial ran in July 1941, and for nearly a decade no one could say how many people saw it (Television Academy). That changed in 1950 when Nielsen came out with the rating point, or estimated percentage of TV households that tuned into a program at the same time (Nielsen). Networks then priced airtime against it, and advertisers built schedules by adding up points, which gave a TV buy a figure that teams could budget around.

In 1964, the industry created the Broadcast Rating Council, later renamed the MRC, to audit the ratings services independently (MRC history; US DOJ). That body went on to set the audit standard for every major channel that followed.

Digital display

Created in 1994Measurable in 2014

Display's first unit counted ads that loaded, not ads that were seen. From the first banner ad, which ran on HotWired in October 1994 (The Guardian), advertisers paid for an impression the moment an ad loaded, whether or not anyone could see it. For two decades, budgets went to ads that could be served below the fold or in tabs no one opened.

The fix arrived in March 2014, after eighteen months of industry debate, when the MRC approved the viewable impression as a unit for buying and selling. Under that definition, an ad counts only when at least half of it is on screen, in the tab the person is looking at, for at least one second (IAB). Ads below the fold count only once someone scrolls to them, and ads in background tabs don't count at all. Written guidelines followed that June (MRC/IAB guidelines).

Search

Created in 1998Measurable in 1998

Search broke the pattern, with a measurement unit from day one. When GoTo.com launched paid search in February 1998, advertisers bid on keywords and paid only when someone clicked (Slate). Google launched AdWords in October 2000, priced by the impression (Google launch announcement), and moved it to pay-per-click in February 2002 (Google pricing announcement). Conversion tracking soon let a marketer follow a search from click to sale and work out what each customer cost.

The catch was that one company controlled every step. Google set the unit, counted it, and reported it across roughly 90% of search (NPR). No industry body defined it, and independent verification trailed by more than a decade. The MRC began accrediting Google's click metrics in 2018 (Marketing Dive) and still accredits its Search clicks and invalid-click filtering (MRC accreditation letter).

Podcasts

Created in 2005Measurable in 2016

Podcasts went mainstream in June 2005, when Apple built them into iTunes (Apple). Advertisers bought ads on download counts that each show measured its own way. With audience data lagging behind other digital media in availability and quality, large brands were slower to invest than direct-response advertisers, which could track sales through promo codes (Digiday).

In September 2016, IAB Tech Lab, the digital ad industry's technical standards group, published the first shared guidelines for counting podcast audiences (IAB Tech Lab guidelines, v1.0). By December 2017 it had defined the download (IAB Tech Lab version history), and a year later it began certifying that podcast companies counted downloads the same way (IAB Tech Lab press release).

Social

Created in 2007Measurable in 2004

Social advertising grew up inside closed platforms. Facebook built the model in November 2007, selling the ads and reporting the results itself (Fortune), and Twitter, LinkedIn, Instagram, Snapchat and Pinterest followed. Unlike earlier channels, social never had to invent a measurement unit. Social sold ads by the impression from the start, using the IAB's 2004 definition: one ad delivered to a screen (IAB guidelines).

Trust broke in September 2016, when Facebook admitted it had overstated how long people watched videos on its platform (The Drum). Advertisers led by Procter & Gamble and the Association of National Advertisers pushed for independent audits (MarTech; Marketing Dive).

The industry now shares an MRC-audited definition of the ad impression, but not the verification behind it. Accreditation is granted product by product and metric by metric. As of November 2026, Meta, Snap, Pinterest and LinkedIn hold it for at least part of what they sell, TikTok relies on outside verifiers, and X and Reddit are not on the list (MRC).

Retail media

Created in 2012Measurable in 2024

Amazon began selling pay-per-click ads on its own site in 2012, letting brands bid on shopper search terms (ProPublica). Within a decade the channel had become a major business. Amazon alone reported $31.2 billion in global ad revenue in 2021 (CNBC), and Walmart, Kroger, Instacart and CVS had built networks of their own (IDC).

Because the retailer owned both the ad space and the checkout, retail media could link an ad to a purchase from the start. Only the retailer could see that link, though. Each network set its own rules for which purchases counted, so brands took every report on trust, with no reliable way to compare them.

In January 2024, the IAB and MRC published shared guidelines that defined the attributed sale: a purchase credited to an ad because the same shopper saw or clicked it within a set window beforehand, often 3 to 30 days (IAB/MRC guidelines). Networks are asked to disclose that window and to say how much of each figure is modeled rather than counted from tracked purchases. Whether an ad caused a sale that wouldn't otherwise have happened, known as incrementality, is measured separately.

The lesson

Every major channel required budget before attribution could be proved. Billboards traded on estimates for decades and television sold ads for nearly ten years before its first ratings. Display, podcasts, social, and retail media all grew while the industry still argued over what to count. Even search, which had its unit from the start, ran for years on figures only the seller could check.

What turned each channel from an experiment into a dependable budget line was a carefully built, widely agreed unit that told the business what its spend would buy.

AI search has reached that same point. The demand is here, but the unit is not, and that is what the AI Search Measurement Standard is working to close.

Chapter 2

You can see the shiftYou can't prove it

2 minute read

With no agreed upon measurement unit for AI search, marketing leaders are in a difficult position. They see buyers using AI platforms to research, but they can’t clearly point to the impact that a budget for AI search will return.

No one has this figured out yet and the questions being asked are consistent across industries and company sizes. The questions below are quoted directly from marketing, content, growth leaders, and founders.

The concerns that are keeping marketing leaders up at night

No revenue number to point to. Marketing leaders can already track how often AI answers mention their brand or cite their content. What they can't say is how a rise in either shows up in pipeline or revenue.

  • How can we track and report on our AI search visibility to justify investment?Content Director, Data Company
  • What number is going to go up? And how do we know that we're getting our money's worth? We want to understand the ROI.Founder, Business Consulting Firm
  • Would we make back the money? That's what I'm trying to figure out.Chief Strategy Officer, Global Consulting Firm

No answer for leadership. When marketers take mention counts, citation rates or AI referral traffic to the executive team, the first follow-up is usually how those numbers were measured, and without a shared standard, that's hard to answer.

  • Trying to get leadership to understand this investment is difficult.Marketing leader, Fintech Company
  • What is the business impact and justification we can take to management?Marketing Leader, Food & Beverage Company
  • How does the platform help customers address pipeline attribution and prove ROI for AI search visibility to executives?Marketing Leader, Hospitality Company

Budgets on hold. Without an answer the executive team will accept, the safest move is to wait, so budgets stay small or sit unallocated until the impact is clearer.

  • We've started to run some ChatGPT ads, but the budget has been pretty low right now because we don't know the actual impact.Marketing Leader, Real Estate Company
  • I don't know the proper way to evaluate this, other than saying ‘Here's hundreds of thousands of dollars, let's go build this, and we'll see’.Director of Marketing Acquisition, Data Infrastructure Company

Tools that disagree. Marketers who go looking for a number often come back with several, each telling a slightly different story about how their brand shows up in AI answers.

  • Every tool on the planet now thinks that they're good at this. And I'm not seeing results.Vice President of Marketing, Business Consulting Firm
  • It seems to be that every AI visibility tool we're using has slightly different results.Senior Director of Marketing, InsurTech Company

Until marketing leaders can answer these questions with confidence, AI search will struggle to earn a line in the budget. A shared unit is what makes that answer possible, and the first step toward one is understanding why today's numbers fall short.

Chapter 3

What your AI search dashboard misses

6 minute read

Marketing teams can cobble together a few signals on AI search, usually in a visibility dashboard that shows where the brand appears in AI answers and the traffic those answers send, pulled from web analytics. Both help make the case for investing, but neither can show what that presence is worth next to the other channels competing for the same budget.

Marketing mix modeling (MMM) is built to make that comparison, but it needs a weekly count of how often buyers came across the brand, the same input TV ratings and search impressions provide. AI Search Impressions are designed to supply that count. Today's dashboards fall short in three ways:

  1. Dashboards don't show full influence. Buyers decide inside the answer, and current tracking can't count that exposure.
  2. AI search visibility numbers wobble. Results move with how a question is phrased, when it runs and which tool asks it.
  3. No line to revenue. Even steady numbers can't tell a marketing leader what AI search drove.

The sections that follow explain each gap, and why closing it takes a shared unit rather than a better dashboard.

1. Dashboards don't show full influence

The most important moments in the buyer's journey now happen inside AI answers, out of marketers' view. Buyers ask ChatGPT, Gemini and Google AI Mode which brands to consider and which one to choose, and the answer shapes that decision before anyone clicks. A brand missing from those answers can lose the sale without knowing it was ever in play.

To see whether they showed up, most teams rely on two signals: prompt tracking and referral traffic.

Prompt tracking runs the questions a brand believes buyers ask and counts how often the brand is mentioned or cited. It shows where the brand appears, but not how many buyers asked those questions, so a mention in a rarely asked prompt counts the same as one in a popular prompt.

Referral traffic counts the visits AI platforms send to a brand's site, and it runs low. Sites aren't always told where a visit came from. A click in ChatGPT's browser version may carry that label while the same click in its desktop app doesn't, so those visits get counted as direct traffic or credited to another channel. And since many buyers get their answer without clicking at all, traffic misses much of the influence anyway.

Both signals are a useful start. Together they can tell leadership how often the brand appeared in AI answers this quarter, but not how many buyers saw it or what it did for pipeline.

“...I don't really know because I don't see it in a dashboard. It's more hearsay.”

Chief Security Officer, Cybersecurity

2. AI search visibility numbers wobble

Even the signals that dashboards do capture don't hold still. The same brand's mention rate can look very different depending on how a question is phrased, when it's asked, and which tool does the measuring.

In an AirOps study tracking which banks were named across 17,820 AI answers about personal checking accounts, the findings really showed how much these factors shift the results. The study covered US English questions on ChatGPT and Gemini only.

The question changes the answer. When people described who they were and why they were asking, for example "I'm a high school graduate opening my first checking account for college," AI answers named the banks being tracked about twice as often. On ChatGPT, the share of answers naming them rose from 20.4% to 41.3%, and Gemini saw a similar jump.

The questions being tested aren't always the ones people ask. Many tracking setups use questions written by software instead of by buyers. When the study ran both kinds on ChatGPT and Gemini, the software-written questions named the tracked banks about as often as the questions people typed in only 2 of 14 cases. Everywhere else, how often a bank appeared depended on which set of questions was used.

Answers change from run to run. AI platforms can give a different answer each time the same question is asked, so a single run proves little. In the study, most questions had to be asked about 30 times before the share of answers naming each bank held steady, and some took 41. The IAB's new guidelines make the same point: "Single-response measurement is not measurement," and results should be reported "as a range rather than a single number" (IAB, Aug 2026).

Methods differ. More than 20 companies now sell AI search visibility metrics, "each using different methodologies that can produce different answers for the same brand or publisher" (IAB).

"It seems to be that every kind of like AI visibility tool we're using has slightly different results."

Senior Director of Marketing, InsurTech

Shifting numbers are normal for a new channel, and they settle once the industry agrees on how to count them. Display went through the same process: when the viewable impression became the trading standard in 2014, readings from 11 accredited vendors still varied, and the Media Rating Council set out to narrow that spread to within 5 to 10 percent (IAB). Every channel's measurement gets messy before it gets reliable.

3. No line to revenue

Even if those numbers held still, the most important question would stay open: what did AI search drive? To answer it, a marketer has to link what buyers saw in AI answers to what they bought. There are three usual ways to make that link, and AI search breaks each one.

Clicks undercount. The simplest link is the click. A buyer clicks through, lands on the site and buys, and the sale is credited to that channel. But many buyers who see a brand in an AI answer never click, because AI platforms are built to answer the question inside the chat and give people little reason to leave. Instead, they search for the brand or type its URL days later, and the sale is credited to search or direct traffic.

Platform data is thin. On other channels, the platforms fill that gap by reporting clicks and the queries behind them. AI platforms share far less, though, like every channel before them, they're opening up more over time. As of this writing, here's what they offer. Google Search Console, Google's free reporting tool for site owners, has a generative AI report that counts impressions from AI Overviews and AI Mode combined, by page, country, device and date. It shows no queries, and clicks from those answers are folded into overall search totals (Google). Bing reports AI citations and a sample of the searches its AI ran to build answers, with no clicks or traffic (Bing). ChatGPT's business profiles count mentions and website clicks, but keep prompts private and count a mention only when an answer cites the brand's website (OpenAI).

Percentages don't add up. That leaves the mix model, which doesn't need clicks at all. It compares how much exposure each channel delivered week by week with how sales moved, then estimates what each channel contributed. But it needs a count, and the best a dashboard offers is a percentage. A 40% share of voice could mean 400 buyers saw the brand or 4 million, depending on how many people asked. That's why Google's own mix modeling tool, Meridian, accepts only counts that can be summed, such as impressions or clicks, and rules out rates and percentages (Google Meridian).

"How do you determine which content opportunities are commercially valuable rather than just generating visibility?"

Founder, Media Company

"...there's this weird disconnect where I see very high share of voice. I see LLM traffic to the website, but there's a disconnect, like something isn't working."

Vice President of Marketing, Data Company

Why it can't wait

Display took twenty years to agree on what counted as a view. AI search won't get that long, because its audience arrived almost overnight. ChatGPT launched in November 2022 (OpenAI), and less than four years later OpenAI reports more than 1.2 billion weekly users (OpenAI, Oct 2026). Google reports more than 2.5 billion monthly users for AI Overviews and more than 1 billion for AI Mode (Google, Aug 2026). The share of website visits these platforms send is still small but climbing fast, from 0.02% in 2024 to 0.32% in early 2026, according to one panel of more than 100,000 sites (SE Ranking, Jun 2026).

The audience is already here. What marketing leaders need now is a way to show what that audience is worth, and the measurement most of them rely on can't give them one.

Chapter 4

What it takes to answer "Is AI search working?"

2 minute read

Every channel that became a dependable budget line got there once marketers had three things in place: data that counted exposure, a methodology to estimate what that exposure drove, and a way to act on the result. Like building blocks, each component uses the one before it.

AI search needs the same three blocks. The AI Search Measurement Standard was formed to build these pieces for AI search. Here's how the group is approaching each one.

DataMethodologyDecisionMeasurement

Data

The AI Search Measurement Standard started with data. The first goal is a metric a methodology can use: AI Search Impression, a weekly estimate of how often buyers see a brand in AI answers. AI platforms share little of their own data, so for now the metric is built from panels and sampled answers, with data providers supplying what people ask AI. It will earn more trust as better data comes out.

Methodology

Once the data becomes a unit, the next step is putting it to work in a measurement framework, which shows what AI search drove alongside everything else in the plan. Of the frameworks marketing teams rely on, the AI Search Measurement Standard chose to start with marketing mix modeling (MMM), because it is generally accepted as the best way to measure non-click-based channels.

No measurement is perfect, but a method the industry agrees on can be checked by anyone and compared across brands. Building that agreement is the goal of the group's measurement partners, brands and data providers.

Decision

Knowing what AI search drove matters only if it shapes the next budget, which is why the third piece of the group's work focuses on decisions.

Once AI search sits in the mix model next to every other channel, marketing leaders can compare its return to other channels. They can also see which topics and pages draw the most exposure, and what they can change in their content to earn more. That gives them a clear answer they can defend to finance and leadership, to the question they started with: should we invest more in AI search, and where?

"One consistent learning from over a decade in marketing measurement is that every emerging channel must mature through the same three milestones: standardizing raw exposure data, calibrating the methodology, and translating results into actionable budget recommendations. A channel only becomes truly investable when the industry agrees on a rigorous, common currency. Our aim is to give the industry the empirical confidence to treat AI search not as an unmeasured experiment, but as an accountable growth engine."

Hitesh WadhwaniCo-founder of Google MeridianGoogle Meridian logo

No single brand or vendor can build all three alone. The next chapter walks through each block, starting with how the AI Search Impression is built.

Chapter 5

Introducing AI Search Impressions, the first piece of the AI Search Measurement Standard methodology

6 minute read

Here's how the AI Search Measurement Standard built the AI Search Impression, including what the group tested, what it found and what it decided along the way.

The impression is built with a three-step framework that works whichever AI search tracking tool a brand uses. AI visibility companies are part of the effort too, including Bluefish. The result is a number that plugs into existing measurement models, starting with MMM, so marketing and finance can weigh AI search against every other channel.

What teams can do with it

  • Evaluate AI search in a marketing mix model, as a weekly series alongside TV, search and everything else in the plan.
  • Start ROI work from a shared number that marketing and finance both accept.

The methodology

Step 1: Topic mapping

Start with topics, then choose prompts.

Most AI search tracking tools start with prompts: a brand writes the questions (prompts) it expects buyers to ask, then tracks how often it appears in the answers. The AI Search Impression starts with topics instead, then chooses prompts within each one.

A topic is a product category a brand competes in within a specific region, described from the market's point of view. For example, "US credit cards" is a topic, while a brand or product name is not. The group calls this approach topic mapping, and it lets marketing teams:

Topic mapping has two advantages, it lets marketing teams:

  • Follow a buyer's entire journey through one product category. This shows where a brand performs well in AI search and where it falls short.
  • Scope tracking to the categories that matter, rather than the whole brand. For brands with several product lines and regions, a brand-level measure can hide where AI search work will have the most impact. One leading corporate card provider had 55% share of voice in AI search overall, but only 1% in answers to the questions its enterprise buyers ask.

Choosing prompts within each topic

Within each topic, the AI Search Impression is built from a set of prompts, so choosing them comes down to two questions.

1. Do people ask this? A prompt should count toward exposure only if buyers ask something like it. To ground prompt lists in demand, the group drew on three sources: questions people sent to AI platforms, supplied by the AI Search Measurement Standard’s data partners; search engine queries; and Google's "People also ask" questions. It then grouped similar questions and removed near-duplicates, aiming to cover each distinct buying question in a topic.

2. Which wording changes the answer? The same buying question can be asked many ways. Tracking every version is costly, and tracking only one can miss what buyers see. So AirOps researchers tested which differences move a brand's results enough to track separately.

Each test held the question, platform and region fixed and changed one element at a time: who is asking, how the question is phrased, how detailed it is and how many times it runs. The study covered eight checking-account questions on ChatGPT and Gemini in US English, and four findings shaped the tracking setup:

  • Who is asking matters, so the setup tracks personas. Adding a buyer audience significantly changed mention rate in 18 of 40 comparisons on both ChatGPT and Gemini. On ChatGPT, a 23.8% baseline ranged from 11.3% to 51.3% depending on the audience.
  • Rewording rarely matters, so it doesn't track many versions of one question. First-person phrasing had no effect on ChatGPT and changed results for 1 of 8 questions on Gemini.
  • Detail matters for some questions, so it tracks a mix. Long-tail prompts, the longer and more specific questions, lifted mention rate from 21% to 40% on ChatGPT and from 14% to 20% on Gemini. The change was significant in 2 of 8 questions on each platform.
  • Answers vary, so every prompt is repeated. AI platforms can give a different answer each time, so each prompt runs 10 times by default.
ExperimentResultWhat it means for tracking
PersonaAdding a persona moved mention rate in 18 of 40 comparisons per platform (ChatGPT: 23.8% baseline, 11–51% with an audience)Tracking personas is significant in some engines
Situational vs direct phrasingFirst-person phrasing had no effect on ChatGPT and changed 1 of 8 questions on Gemini.Phrasing style matters little, so there's no need to track many rewordings of one question.
Run countAbout 95% of results settled within 10 runs (median 9–14)Run each prompt 10 times by default.
Short vs. long tailLong-tail prompts lifted ChatGPT mention from 21% to 40% and Gemini from 14% to 20%, significant in 2 of 8 questions per platformTrack a mix of short and long-tail prompts

Defining topics first, building prompts from buyer-query data and letting the experiments set the tracking rules kept every number tied to the market being measured.

Step 2: Prompt volume

Estimate how often people ask each prompt.

Tracking tools give every prompt the same weight, however often people ask it. A question asked thousands of times a week should count for more than one asked a few times a month. Google publishes search volume for keywords, but AI platforms publish nothing comparable. So the AI Search Measurement Standard estimated demand under two rules: ground the estimate in observed data wherever possible, and keep the method open for anyone to check.

The first version uses panel data. Research panels are groups of people who opt in to share how they use apps and websites, including AI platforms. The source is an opt-in panel that includes ChatGPT messages. The method uses public benchmarks such as OpenAI Signals to correct for groups the panel over- or under-represents.

The method has five steps:

  1. For each tracked prompt, find panel messages that ask the same question, even in different words, then run a separate check that the intent matches.
  2. Assign each message to its single closest prompt so no demand is counted twice.
  3. Count the matches, correct for matching errors and combine panels.
  4. Weight the results to match the user population, then scale them up using public figures on message volume to get a monthly count.
  5. Split each monthly estimate into weeks so demand can be reported weekly.

The estimate will sharpen as AI platforms share more data about what people ask. The group is also testing other ways to estimate prompt volume that may be folded into the method.

Step 3: AI Search Impressions

Count how often buyers see the brand.

Topic mapping sets which topics and prompts to track, and prompt volume estimates how often people ask them. The final step counts how many times buyers see the brand in answers to the tracked prompts.

What counts as exposure

First the group had to decide which appearances should count. A mention or citation shows that a brand appeared in an answer, but not that the answer recommended it. Today's tracking scores a brand listed tenth of ten the same as the brand the answer recommends. It also can't separate a brand that appears in every answer without being chosen, or one mentioned mostly in a negative light. So it built a composite score driven mainly by recommendation rate, meaning how often the AI selects a brand as the option to go with.

What the group weighed: mentions and citations.

What it decided: a composite score driven by recommendation rate, because a brand can be mentioned everywhere and still never be the one chosen.

How the count comes together

For each prompt, the metric multiplies the prompt's estimated volume by the brand's composite score. Recommendation rate drives that score. Positive mentions earn extra credit, and other mentions earn less. Weighting by volume means the questions buyers ask most count the most. The result is a modeled count of how many times buyers see the brand recommended in AI search.

Why exposure, not clicks

Counting exposure rather than clicks is deliberate. Many buyers who see a brand in an AI answer never click, and return later by searching for the brand or typing its URL. Privacy rules and browser changes are also making click-based attribution harder, so more teams are turning to marketing mix modeling, which needs a count of exposure to work.

What it measures, and what it doesn't

AI Search Impressions estimate exposure, and while they don't prove ROI or revenue impact on their own, they're a key input to that work. Getting to revenue impact requires testing the metric with brands inside their mix models, which is where Paramark and other measurement partners who are part of the AI Search Measurement Standard comes in.

The full documentation, with every assumption and calculation, is coming soon.

Get involved

Join the AI Search Measurement Standard and help shape it. Leave your work email and we will be in touch.

Chapter 6

What's still open in AI search impressions and where the initiative goes next

1 minute read

A shared unit for AI search has to be tested and refined in practice. That's why the AI Search Measurement Standard is publishing its method in the open, so that anyone can examine the assumptions, test the results and suggest improvements.

Here’s what’s next in this work:

More data will sharpen the estimate. AI Search Impressions depend on knowing how often people ask each question. AI platforms share little of that data, so for now it comes from opt-in panels, and the AI Search Measurement Standard reports the number as a range to avoid false precision. Google gave brands more data over time through tools like Search Console, and the group hopes AI platforms will do the same, with each new source sharpening the estimate.

Building for more measurement frameworks and platforms. AI search impressions work today in marketing mix modeling, which shows how much a channel added over a quarter. They can't yet show which campaign led to a sale, or whether the sale would have happened anyway. Versions for different marketing measurement methodologies are next.

Reach and frequency aren't separated yet. AI Search Impressions count total exposure. They can't yet tell whether 1,000 impressions means 1,000 people seeing a brand once or 100 people seeing it ten times. Separating how many people saw a brand (reach) from how often they saw it (frequency) is still an open question.

Prompt volume estimates will add more AI platforms and categories. For now, the estimate of how often people ask each question comes from ChatGPT data. Other AI platforms answer buying questions too, and later versions will cover them. Testing the method across more brands and industries will also make it more useful for companies outside the group that built it. Both depend on data partners willing to contribute.

Join us in setting the standard

How to get involved

  1. Join the AI Search Measurement Standard. Become a brand, data provider, or measurement leader building the standard, and help steer what it tests next. Members of the group can review, test, and extend each version as it's released.
  2. Follow the work. Sign up to get each new version of the method as it's published.

The first version of the metric is almost ready. The rest of the standard will be built by the people who join now. Join now.

Webinar

From Black Box To Budget:Inside AI Search Measurement

Friday, Nov 20 · 2:00 PM ET / 11:00 AM PT

  • Hitesh Wadhwani

    Co-Founder of Google Meridian

  • Pranav Piyush

    CEO
    Paramark

  • Matt Hammel

    Co-founder & COO
    AirOps

  1. 1Join the webinar

    See where the standard stands today on Nov 20.

  2. 2Follow the progress

    Sign up and we’ll keep you posted as the group makes progress.

  3. 3Help shape the method

    Tell us who you are and how you’d like to take part.