Editor’s Pick Revenue Editors Picks

ChatGPT gives publishers’ AI sales pitch a new test

Ivan Massow 11 Nov 2026 · 6 min read

An analyst in red holds an enormous magnifying glass in front of a client in blue. The lens is clogged with speech bubbles, blocking the client’s view. The caption reads: Everything is much clearer now.

Cartoon · AI-generated

OpenAI has begun supplying a number that publishers selling AI visibility services could previously only estimate from outside. Its Business Hub, launched in beta on 10 November, gives verified businesses information about appearances and website clicks in ChatGPT. For publishers such as Future and the Daily Mail, the change is a test of what their commercial advice is worth. Advertisers gain another way to challenge a campaign report. Suppliers charging for a count face a harder sell. The opportunity is to explain and improve business outcomes, rather than simply report that a brand appeared in more answers.

This is a substantial change from the reports of an unreleased dashboard circulating last week. OpenAI has now announced the product. It combines business information, aggregated insights and, for eligible US ecommerce and online service businesses, a separately labelled conversational agent. The broader business-claiming offer and the narrower US agent rollout should not be confused.

The useful detail is the definition of a mention. For an online business, OpenAI’s help page says it counts a reference to the business’s website, rather than the brand’s name alone. Discovery topics summarise themes; they do not reveal individual prompts or directly measure customer intent. A blank metric can mean insufficient activity or unavailable reporting, rather than zero.

Farhad Divecha, group chief executive of marketing agency Accuracast, identified the commercial pressure in Krystal Scanlon’s 10 November Digiday report. “If their estimation data was wildly erroneous, that will make us rethink paying for those tools,” he said. That is an agency buyer’s conditional judgement, not evidence that any named supplier’s estimates have failed.

Advertisement · Noah

Publishers have already found customers for services known as generative engine optimisation, or GEO: work intended to help a business appear in AI-generated answers. Sara Guaglione reported in Digiday on 1 October that Future had more than 30 clients and renewals. Ziff Davis also reported clients. These are signs of demand, not audited figures for the size or profitability of the market.

Future’s chief revenue officer Mike Peralta described the proposition at the Digiday Publishing Summit, as quoted in that report: “You want to make it simple for the brand to understand and buy”. The task now is to make the evidence equally intelligible. A package can be easy to purchase while leaving the purchaser unclear about what improved.

The Daily Mail’s experiment shows the distinction

Jessica Davies reported in Digiday on 26 October that the Daily Mail had 12 to 15 tests running across the UK, US and Australia, ahead of a planned product early next year. Its team selected relevant prompts, observed answers and citations over time, then created branded content where it saw an opening.

In one test for an unnamed personal-finance brand, director of SEO and editorial ecommerce Carly Steven reported visibility rising from about 50% to nearly 75% across target prompts. That is a useful experimental result to investigate. It is not a measure of the share of all potential customers reached, and it does not establish additional sales.

Those percentages describe appearances in a chosen set of questions. They are not forecast probabilities. The commercial question is whether that set resembles questions customers actually ask, and whether the apparent improvement survives repeated observation. The publisher’s role in choosing the test makes disclosure of the denominator particularly valuable.

An explicitly hypothetical example makes the gap clear. Suppose a campaign increases a brand’s appearances from 50 to 75 of 100 monitored answers. That is 25 additional appearances in the test, not 25 additional shoppers. If the underlying questions are rarely asked, the result may have little commercial weight. If they concern expensive purchases and closely match real demand, even a modest gain could matter.

OpenAI’s figures can provide a different view of activity, but the two series should not be expected to match. A publisher’s fixed prompt panel is an experiment it designed. Platform counts describe activity under the platform’s definitions and reporting limits. A discrepancy is a reason to investigate the methods before declaring either a success or a failure.

The defensible service is interpretation

The new dashboard does not eliminate the work publishers can sell. It does make an undifferentiated tracking service easier to question. A specialist title can bring product knowledge, original testing, useful explanations and an understanding of the language customers use. The value lies in those capabilities, not privileged access to an answer engine’s approval.

That is also why a publisher should separate commercial content from independent evaluation. A recommendation earns its usefulness by helping a reader decide, including deciding against a purchase. A client paying to improve its representation is buying a different service. Blurring those two propositions would spend the editorial authority on which the commercial offer depends.

There is a practical distinction between making information accurate and purchasing influence. OpenAI says the Business Agent conversation is labelled and separate from its independent answers. Sponsored placements are separate again. Those boundaries give a publisher a useful test for its own pitch: does the client understand whether it is buying content, consultancy, distribution or advertising?

Nor does completing the information supplied to a platform guarantee a recommendation. OpenAI’s setup documentation makes that limit explicit for product feeds. A publisher can reasonably offer better structured information and test what follows. It cannot turn the platform’s discretion into a promised position in an answer.

A second limit appears in the data terms

Publishers considering a managed service should read beyond the product announcement. OpenAI’s Businesses in ChatGPT terms, updated 9 November, say OpenAI owns Insight Data. They permit its use to measure, analyse, optimise and report on use of the services, while restricting disclosure and commercial exploitation unless OpenAI authorises otherwise in writing.

The terms allow access by service providers acting on the business’s behalf for the permitted purposes, subject to equivalent restrictions. That is materially different from a general right to resell the numbers or build a shared commercial dataset from them. A consultancy report for one client and a pooled benchmarking product are different propositions under this wording.

This matters to the economics of a publisher’s offer. Comparing results across clients could make advice more useful and cheaper to produce. But a publisher cannot assume that access granted for one client’s reporting supplies unrestricted material for its next product. The data permissions need to fit the proposed service before the rate card is built around it.

The same distinction helps prevent an overreaction. Access restrictions do not make the dashboard worthless. A client’s own measurements can still help assess its work with a service provider. They do mean the business model should be explicit about which evidence belongs to the client, which comes from independent research and which can be reused.

Sell the improvement, not the dashboard

The strongest offer would connect three observations without pretending they are interchangeable: how content performs in a repeatable test, what the platform reports about the business, and what the client records after a visit. An answer that creates awareness without a click may have value. Establishing that value requires evidence beyond labelling an appearance a conversion.

A publisher could test a campaign against comparable untreated content or a pre-agreed baseline, while recording other changes that might affect demand. That will not produce perfect attribution. It would, however, give a client a better basis for renewing than a chart whose prompts, observation frequency and scope change whenever the result disappoints.

Cost matters alongside uplift. Research, content production, specialist review and client reporting all take time. Neither a growing client list nor a reported visibility gain reveals the margin left after that work. Publishers should avoid building a labour-intensive consultancy and describing it to themselves as effortless new inventory.

Business Hub makes useful AI adoption easier to examine. It also returns the measurement advantage to a platform that defines the available data. The sensible response is neither to abandon publisher expertise nor to dress ordinary branded content in a new acronym. Future, the Daily Mail and their peers have a credible business to develop if they can demonstrate what their intervention changed. OpenAI has just made a count alone a less convincing answer.

Sources

Source: Tomorrow’s Publisher