Where AI search goes next: who sells the answer, who measures it, and how AI agents choose

What did we find?

1. The platforms shipped more ads than measurement. In twelve months OpenAI went from ad principles to a $1 billion run rate and a long list of measurement partners for advertisers. Google added ad formats inside AI Mode and checkout on its own pages. Organic visibility got an impressions-only report from Google, a citations report from Bing, and a referral tag from OpenAI.
2. Each platform reports the surface it sells and withholds the surface it gives away. Where regulation or payment forces disclosure, data appears. Elsewhere it stays vague.
3. Cheap AI visibility trackers are the most exposed; cross-engine comparison is the most durable. Every major AI platform’s terms prohibit the automated collection most trackers use, and suites are already buying the vendors.
4. Paid will gain share, but not because organic is uniquely hidden. Google has withheld paid data too, and AI answers cut paid clicks as well as organic ones. Budgets drift toward what can be measured.
5. Measurement improves where someone is forced. UK sites should see Google AI clicks by mid-2027. ChatGPT, Gemini, Claude and Perplexity stay black boxes through 2027.
6. AI agents are not the rational shoppers the pitch decks describe. They choose by spec when one product clearly wins. Otherwise list position, badges, brand recognition and well-written copy decide, and the weights change between model versions.
7. Brand still matters, in a specific way. Being familiar decides whether an AI names you at all. Verifiable detail decides how you are cited and whether an agent can pick you.

What have the platforms done so far?

In twelve months the platforms shipped more monetization than measurement, and the measurement they shipped stops at their own borders.

What the measurement products leave out. Google’s generative AI report shows impressions by page, country, device and date, with AI Overviews and AI Mode combined. It has no clicks, no queries and no API. Google Ads reports ads inside AI Overviews as “Top Ads,” says it “currently doesn’t offer segmented reporting,” and gives advertisers no way to opt out of that placement. Bing shows citations and the search phrases its AI used, but no clicks. OpenAI’s advertisers get a Conversions API, a pixel and partner attribution; for organic brand visibility inside ChatGPT we found no OpenAI product at all. Google Analytics 4’s new AI Assistant channel “excludes Google’s AI Overviews and AI Mode.”

The money at stake. Alphabet’s Search & other revenue was $63.3 billion in the second quarter of 2026, up 17%, after 19% in the first quarter. Paid clicks rose 13% in both quarters, according to Alphabet’s 10-Q filings. Sundar Pichai says AI Mode passed a billion monthly users and is “driving an incremental increase in Search queries overall.” OpenAI told investors to expect $2.5 billion in ad revenue in 2026 and $100 billion by 2030, according to Axios; Digiday estimates about $330 million was actually booked through August. Microsoft’s search advertising growth slowed from 21% to 10% over its fiscal year. Amazon’s ad revenue grew 24% to $17.2 billion in the first quarter, and Andy Jassy says “Advertising will do well in a world of agentic commerce.”

What will the platforms do next?

The incentives are written down in filings, help pages and regulatory submissions. What follows is our reading of them.

Google: sell the answer, report the impression. Google has merged AI Overviews and AI Mode into “one seamless Search experience.” It told the UK regulator that splitting data by feature “would not accurately reflect user behaviour,” and the regulator accepted that. So the data due in December will most likely be one bucket covering both features, for UK users only, with a carve-out for commercial information. The ad side moves faster, with new AI Mode formats and checkout on Google’s own pages. Google already calculates share of voice in AI surfaces against named competitors for merchants in five countries. That it does so for shopping and not for publishers is a choice, and the likeliest next step is that kind of competitive view as a paid or merchant product rather than a free Search Console report.

OpenAI: measurement for buyers, a referral tag for everyone else. The advertiser tools came first and fast. Organic visibility has a referral tag. OpenAI’s own forecast needs ad revenue to grow about 40 times by 2030, which argues for more formats and markets, not a free dashboard that would show brands how to appear without paying.

Microsoft: generosity as distribution. Bing has the least to lose and gives the most data. Every marketer who learns to read its report is a marketer who optimizes for Copilot. Expect more features and perhaps an API, but not clicks; Bing says its citation share “does not expose competitor domains, represent traffic share.”

Amazon: own the shopper, sell the prompt. Rufus usage is growing, sponsored prompts are live, and Amazon has fought outside shopping agents in court. For brands it is the most measurable AI surface, because it is an ad product.

Perplexity and Anthropic: no ads, so no ad data. Perplexity’s ads earned $20,000 against $34 million in 2024 revenue, and advertisers told Adweek they could not measure them. Its executives now say a user “needs to believe this is the best possible answer.” Anthropic has pledged that Claude stays ad-free. Neither offers brands a visibility product, and neither has a commercial reason to.

Will AI visibility tools be pushed out by the platforms?

The vendors have a legal problem before they have a competitive one. OpenAI’s consumer terms forbid users to “Automatically or programmatically extract data or Output.” Anthropic’s forbid automated access except “via an Anthropic API Key.” Perplexity’s ban scrapers used “to monitor, extract, copy or collect information.” Google’s spam policy names “scraping results for rank-checking purposes.” Yet the leading trackers collect answers from the apps themselves, because API answers differ from what users see. One vendor, Peec, collects “through browser automation rather than official APIs,” according to the analyst firm Sacra.

We found no enforcement against an AI visibility vendor. But Google has shown the escalation path in classic search. First a policy. Then, in January 2025, a JavaScript requirement that broke rank trackers. Then, in December 2025, a lawsuit against the scraping service SerpApi over “hundreds of millions” of automated queries a day. When OpenAI changed ChatGPT in March 2026, a hidden data field disappeared and tools built on it “lost their primary data source overnight.”

The squeeze so far comes from software suites, not platforms. Adobe bought Semrush for about $1.9 billion, HubSpot bought XFunnel, and Sitecore bought Scrunch for about $225 million. Ahrefs gives prompt tracking free with any paid plan. Meanwhile Profound raised money at a $1.8 billion valuation without disclosing revenue. The Interactive Advertising Bureau counts “more than 20 companies” selling these tools, “each using different methodologies and producing different answers for the same brand,” and says only 16% of brands track AI visibility systematically.

Our reading. Monitoring-only tools at $29 to $500 a month are the most exposed: Google, Bing and Cloudflare now give away metrics for their own surfaces, and suites bundle the basics. That is the Google Analytics pattern. The view no platform will build, a brand’s standing across rival AI engines against named competitors, is the same gap that kept rank trackers alive beside Search Console. It should keep the better-funded vendors alive, most likely until a suite buys them. The tail risk is access. If OpenAI or Google applies Google’s SerpApi approach to chat apps, interface-scraping vendors face a shock no amount of funding prevents. Nothing we found shows that is planned.

Will paid ads gain, and do platforms keep organic data vague on purpose?

Paid will gain share. The case that platforms keep organic metrics vague in order to sell ads is weaker than it looks, because they have withheld data from advertisers too.

The precedent that fits. When Google encrypted searches in October 2011, it stopped passing organic keywords to sites, while for ad clicks the query would still flow “to enable advertisers to measure the effectiveness of their campaigns.” By September 2013 a tracker of 60 sites showed “almost 75% of terms being withheld.” Danny Sullivan wrote at the time that terms had “been withheld all along in part to create new Google advertisers.” That is commentary, not proof of intent, and we found no study that measured a budget shift.

The precedents that cut against it. In September 2020 Google cut the paid search terms report, hiding queries behind about 28% of spend for one agency’s clients. In the antitrust trial Judge Mehta found that “Google’s own records show that the privacy rationale was suspect,” and called Google “a monopolist unconcerned about product changes that have decreased advertisers’ autonomy.” Today Google reports ads inside AI Overviews only as “Top Ads.” Advertisers are not the favored party. They are another party with less leverage than Google.

The economics. The incentive is well modeled. A 2008 paper found a monopoly search engine “has incentive not to offer maximal search quality, even when it would be technically costless.” A 2026 Economic Journal paper summarized by Cornell finds platforms “make unpaid listings less informative, so consumers either buy the sponsored result or spend more time searching. Either way, the platform wins.” The FTC alleges Amazon is “replacing relevant, organic search results with paid advertisements”; on Amazon, four of the first twenty products a shopper sees are organic. But there is a limit: a 2026 field experiment found that more ad slots raised revenue by up to 43% and cut conversions by up to 5%.

What is happening to paid now. Google Search revenue is growing faster than before AI Overviews. Third-party data paints a different picture of clicks: one agency, Seer, found paid click-through on informational queries with an AI Overview “still down 68% from baseline,” and a reported Skai benchmark had spend up 17% with clicks up 2%. AI answers compress paid clicks too. Advertisers appear to be paying more per click rather than getting more clicks.

The limits on ads in answers. 63% of US adults told Ipsos that ads in AI search would reduce their trust. Yet 83% of answer-engine users told Forrester they would keep using free ad-supported tiers, and 6% would pay to avoid ads.

Healthcare is a special case. Google’s help page still says it doesn’t show ads inside AI Overviews “for sensitive verticals like adult, alcohol, gambling, finance, healthcare, politics, and more,” and has said so in every archived version since 2025. The one Google statement that says otherwise is a LinkedIn comment from Google’s Ads Liaison, Ginny Marvin, around Marketing Live in May 2026: “Yes we are beginning a small test of ads in AI Mode for the healthcare vertical (in the US for English queries).” She added that Performance Max, AI Max, Shopping and broad match “are eligible to serve ads in AI Mode (and AI Overviews),” and that the first iteration “is limited to creatives that don’t use pinning or text disclaimers.” Her test sentence names AI Mode only. We found no Google source saying healthcare ads may appear inside AI Overviews. Ads above and below AI Overviews have always been open to healthcare, and all of these placements are reported as “Top Ads,” so an advertiser cannot tell which they got. OpenAI keeps ChatGPT ads away from “health, mental health or politics.”

Our verdict. Paid gains through measurability bias. Gartner’s CMO researcher warns that AI-driven budgeting pulls money toward “the channels and stages of the journey that are easiest to tune.” That needs no conspiracy. The stronger claim fits the incentives but is not demonstrated, because the same platforms cut paid data when they could. Vagueness serves platform leverage in general. Paid is where that leverage is turned into revenue, and regulators are the only force pushing the other way.

Will measurement improve?

Yes, unevenly, and mostly where someone is forced to provide it.

Google’s own position in the regulatory record is that requiring click-quality metrics from Google alone “would create an uneven competitive environment, obliging Google to disclose its methodologies while its competitors have no such obligation.” By mid-2027 a UK site should be able to see AI-feature clicks. Whether a US or Swedish site can depends on whether Google builds the feature once and ships it everywhere, as it did with impressions. ChatGPT, Gemini, Claude and Perplexity remain black boxes beyond referral tags and sampled prompts through 2027 and probably 2028.

Do AI agents choose rationally?

By specification when a specification settles it. By brand, position and badges when it doesn’t, and in a real catalog it often doesn’t. Our separate review of 26 claims about AI shopping agents covers adoption and platform fees; here we look at how the agents decide.

Where they are rational. In the Columbia and Yale ACES experiments, the newest models tested in December 2025 “almost never make any mistake” when one product clearly beats the others on price or rating. Their price sensitivity is “broadly in the same range as” human shoppers’. In a 2026 preprint, a known brand with identical specs beat nine fictional brands in 670 of 670 trials. But once a fictional rival had a slightly better rating, price or review count, the known brand won only 1.7% to 4.6% of the time. Brand works as a tiebreaker, not an override. A 2023 PNAS study found GPT scored higher than human subjects on economic rationality.

Where they are not. When products are close or the request is open-ended, other things decide.

• Position. Moving a product from the bottom right to the top row “leads to a 5-fold increase in selection rate” for Claude Sonnet 4. Telling the agent to ignore position “does not significantly mitigate position effects.”
• Badges. An “Overall Pick” badge raised a 10% baseline to 24.3%, 19.9% and 42.6% across three models. A “Sponsored” label lowered selection.
• Speed. In Microsoft’s simulated marketplace, the first proposal an agent received was chosen 60% to 100% of the time, “a 10-30 fold advantage.”
• Fake social proof. “Sold 50k+” claims cut agent accuracy by 18.5% to 24.5%. Agents fell for a single dark pattern 41% of the time.
• Brand as prior knowledge. GPT-4 Turbo and Llama 3 are “heavily influenced by their latent knowledge of product names.” Across six current models, which brands get recommended tracked “Google search interest as the strongest signal,” not brand-equity scores.
• AI-written copy. AI selectors chose the AI-written product description 89% of the time; human evaluators chose it 36% of the time (PNAS, 2025).
• Version changes. One product’s market share went from 45% to 77% after a single Claude upgrade, and from 25% to 6% after a GPT upgrade.

Human-like, but not human. All 30 cognitive biases tested appeared in at least some of 20 models. More reasoning did not make models less biased. Persuasion techniques raised compliance in 2026 models from 35.3% to 51.3%. But some quirks run the other way from people’s, and personality-survey answers drift toward random once answer order is shuffled. The better word is not emotion but salience: models weight what is prominent, recent, recognized and well written.

Almost all of this comes from simulated stores and API calls. Live consumer products have not been audited.

Do people choose a brand first and justify it with specs later?

For low-involvement and emotionally loaded purchases, the evidence says yes. Where buyers are expert or prices are transparent, it reverses.

People cannot report what drove their choice. In a 1977 experiment, shoppers chose the right-most of four identical pairs of stockings “by a factor of almost four to one,” and “no subject ever mentioned spontaneously the position of the article.” In the choice-blindness experiments published in Science in 2005, only 13% of secretly swapped choices were noticed at the time, and people gave reasons for choices they had not made.

Brand cues change the experience itself. In a 2004 brain-imaging study, “brand knowledge for one of the drinks had a dramatic influence on expressed behavioral preferences and on the measured brain responses.” The same wine labeled $90 instead of $10 was rated more pleasant (PNAS, 2008, 20 subjects).

Preference is built over time. The IPA effectiveness databank of 996 campaigns sums it up on one slide: rational messaging has a “big direct effect, but decays quickly”; emotional priming has a “smaller effect on sales” but “decays slowly.” In business-to-business marketing, Binet and Field found that “differentiation strategies are not very effective as top-of-funnel strategies.”

Buyers mostly buy what they had in mind. In 6sense’s survey of nearly 4,000 business buyers, buyers chose from their day-one shortlist “95% of the time,” and the favorite before first contact won “roughly 80% of the time.” Bain wrote in 2026 that “around 90% of buyers purchase from their Day 1 list” and that “brand reputation, trust, and perceived risk often outweigh product features.” All three of these are industry studies.

Where it reverses. Pharmacists choose national-brand headache remedies “only 9 percent of the time, compared to 26 percent of the time for the average consumer” (Quarterly Journal of Economics, 2015). About 55% of UK motor insurance is sold through price-comparison sites, where “price is a factor of primary importance.” In Google’s own experiment, a made-up cereal brand with no prior exposure won 28% of shopper preference when given reviews, offers and “short descriptions of key product specifications.”

Where AI comes in. In SparkToro’s test, Bose, Sony, Sennheiser and Apple appeared in 55% to 77% of 994 AI answers about headphones, even though the exact list repeated less than once in a hundred runs. Being recommended moves people: in Scrunch’s panel, searches for a brand rose 4.3 percentage points among users with no prior contact after an AI named it. Similarweb found that recommended brands were 2.5 times more likely to be visited within a week. A conversational shopping assistant that subtly steered choices switched 36% of them. About half of consumers check AI recommendations against retailer reviews.

What this supports. The idea that you build a brand with marketing, product and actions rather than spec sheets holds as a statement about building preference over years. It does not make specs useless. Product and differentiation messages belong in the final step before purchase. Agents punish brands that lose on a measurable attribute. And content with statistics gains visibility in AI answers. Familiarity decides whether you are named at all; structured, verifiable detail decides how you are cited and whether an agent can choose you.

What should marketing teams do?

1. Treat AI answers as an awareness channel with partial measurement. The value shows up as later brand searches and direct visits, not tracked clicks. Report reach, share of answers and lift in brand searches, the way broadcast is reported.
2. Get ready for December 3, 2026. Have the Search Console export, the branded-query filter and analytics channel definitions in place so the first weeks of click data are usable.
3. Buy vendor tools only for the view platforms won’t provide. That means comparison across AI engines against named competitors. Prefer vendors that publish their sampling method and run each prompt many times, and treat single-run rankings as noise.
4. Plan for the tool to break. Keep raw answers, prompts and dates in your own storage so a vendor outage does not erase the history.
5. Optimize for what agents reward. Be recognized, keep dates fresh, keep product attributes clean and structured, win on at least one measurable spec, and earn the platforms’ own badges and listings. Never fake social proof.
6. Keep spending on the brand. AI answers favor the already familiar. A brand nobody has heard of does not get named.
7. Expect paid to take share, and test before you follow it. If paid is the only measurable AI surface, budgets will drift there whether or not it performs. Measure the extra sales it adds before shifting money. US healthcare advertisers can enter Google’s AI Mode test through broad match and AI-driven campaigns, without pinned assets or text disclaimers for now.
8. Say in writing what cannot be measured. The influence of AI answers that are never clicked, the chat apps’ prompt volumes and the stability of “AI sentiment” are not measurable with today’s tools.

Limitations

• Many figures come from parties with an interest. These include vendors (Similarweb, Profound, Scrunch, Seer, Skai, Cloudflare, Adthena) and buyer surveys from vendors (6sense, Bain, CEB). Forecasts come from EMARKETER, which sells an AI visibility index. Platforms report their own numbers, such as Google’s “billions of clicks.” Anthropic, whose Claude models Copy for Originals uses in its research, is quoted on its no-ads pledge.
• The agent research is mostly unreviewed preprints from 2025 and 2026, run in simulated stores. Results go stale as models are replaced, which the authors say themselves.
• The classic psychology studies are small or old, and the IPA databank is made of award entries.
• Likelihoods are our judgment from dated evidence.
• Some claims we left out because we could not verify them at source. These include an Apple brand premium, OpenAI’s ad prices, and a measured budget shift caused by “not provided.”

FAQ

Will ads take over AI search answers?

Ads are arriving, but not everywhere. ChatGPT ads reached a $1 billion annualized run rate in under 200 days (OpenAI, quoted by Digiday, August 31, 2026), and Google is testing new ad formats inside AI Mode. Perplexity wound its ads down and Anthropic has pledged that Claude will stay ad-free. Surveys point to a limit: 63% of US adults told Ipsos that ads in AI search would reduce their trust, though 83% of answer-engine users told Forrester they would keep using free ad-supported tiers.

Will third-party AI visibility tools survive?

Probably the ones that compare a brand across rival AI engines, and probably not the cheap single-engine monitors. Google, Bing and Cloudflare now give away metrics for their own surfaces, and suites are bundling the basics (Adobe bought Semrush, HubSpot bought XFunnel, Sitecore bought Scrunch). The bigger risk is access: OpenAI's, Anthropic's and Perplexity's terms prohibit the automated collection most trackers rely on, and Google has already gone from policy to a technical block to a lawsuit against scraping of its classic search results.

Do AI shopping agents choose by brand or by specifications?

By specification when one product clearly wins, and by brand, list position and platform badges when it does not. In a 2026 preprint a recognized brand beat fictional brands with identical specs in 670 of 670 trials, but lost almost every time once a fictional rival had a slightly better rating or price. In the Columbia and Yale ACES experiments, moving a product to the top row raised its selection about fivefold for one model, and an Overall Pick badge raised a 10% baseline to between 20% and 43%.

Can healthcare advertisers show ads inside Google's AI answers?

In AI Mode, on a test basis in the US. Google's Ads Liaison confirmed on LinkedIn in May 2026 a small test of healthcare ads in AI Mode for English queries, open to Performance Max, AI Max, Shopping and broad match, with the first iteration limited to creatives without pinning or text disclaimers. Google's help page still lists healthcare among the verticals excluded from ads inside AI Overviews, and no Google source we found says otherwise. Ads above and below AI Overviews have always been open to healthcare. ChatGPT excludes health from its ads.

Sources

All read on October 16, 2026. OpenAI’s pages refuse automated reading and were read through an archive copy.

• Google: Search Generative AI performance reports; new controls for website owners; About ads and AI Overviews; Merchant Center AI performance insights; New ad formats built with Gemini (May 20, 2026); Alphabet Q2 2026 remarks; Q2 2026 earnings exhibit; GA4 channel definitions; search terms update, 2021; Making search more secure (2011); spam policies; Liz Reid on clicks (2025)
• Healthcare ads: Search Engine Roundtable and PPC Land (June 1, 2026), quoting Ginny Marvin on LinkedIn; Adthena, Pharma in Google AI Mode (vendor)
• OpenAI: our approach to advertising; testing ads in ChatGPT; new ways to buy ChatGPT ads; new ad format and measurement; Digiday, $1 billion run rate; Axios, $100 billion target
• Microsoft: Copilot ads; Bing AI Performance; June 2026 additions; Topics Portfolio; FY26 Q4 metrics
• Others: Anthropic, Claude is a space to think and consumer terms; Perplexity, terms of service; MacRumors, Perplexity abandons ads; Yahoo Finance/Adweek, Perplexity pauses ads; Modern Retail, Rufus
• Regulation: CMA final decision, interpretative notes and summary; European Commission, DMA data sharing and DMA fine; US final judgment, status report and 2024 findings; Penske and Chegg dismissal; FTC, Amazon complaint (allegation)
• Vendors and history: IAB, Measuring Visibility in the AI Era; AdExchanger, on the IAB guidance; Search Engine Journal, the shortcut behind some AI optimization tools; IPWatchdog, Google v. SerpApi; Sacra, Peec; Sitecore, Scrunch acquisition and TechTarget on its price; HubSpot, XFunnel; Semrush, Adobe acquisition; TNW, Profound; MediaPost, the death of Urchin; GOV.UK, not provided; Seer, search terms and AI Overview CTR; Marketplace Pulse, Amazon organic results; White, search engine quality (2008); Cornell, on Janssen et al.; ad load field experiment; Forrester, consumers on ads in ChatGPT; Search Engine Journal, on the Ipsos survey
• AI agent research: ACES; Magentic Marketplace; Incumbent Advantage; brand retrieval and ranking; ranking manipulation; BiasRecBench; dark patterns; AI-AI bias; 30 biases in 20 models; reasoning models and bias; Wharton, persuading LLMs; economic rationality of GPT; survey artifacts
• Brand psychology: Nisbett and Wilson, 1977 (PDF); Johansson et al., 2005 (PDF); McClure et al., 2004, Neuron (doi:10.1016/j.neuron.2004.09.019); Plassmann et al., 2008 (PMC); Binet and Field, The Long and the Short of It (IPA); 6sense, 2025 Buyer Experience Report; Bain, Likelihood to Buy; Bronnenberg et al., 2015, Quarterly Journal of Economics (doi:10.1093/qje/qjv024); CMA, digital comparison tools; Think with Google, the messy middle; SparkToro, AI inconsistency; From Prompt to Purchase (Scrunch); Similarweb, downstream impact; steering by conversational AI; Croud, Consumer Index

How to cite: Copy for Originals, “Where AI search goes next: who sells the answer, who measures it, and how AI agents choose,” October 16, 2026, https://tools.copyfororiginals.com/where-ai-search-goes-next/.