The Ramp AI Index: How Card Data Became a Research Report the Press Cites

Ramp's AI Index says 55.9% of US businesses paid for an AI product in September 2026. The US Census Bureau's own survey put business AI use at 19.8% in May. That 30-point gap is the whole marketing case for a proprietary-data research report: Ramp owns a number nobody else can produce, publishes it monthly, and lets the press do the distribution.

What is the Ramp AI Index, in one paragraph? It is a monthly research report from Ramp Economics Lab that measures the share of US businesses on Ramp paying for AI products, overall and by size, industry and vendor. It is built from card, invoice and bill-pay transactions, not surveys, and it is used by press, analysts and the Fed's FRED database as a business adoption benchmark.

What the Ramp AI Index is, and why a marketer should study it

Ramp sells corporate cards, bill pay and spend software. It is not a research firm. Its economics team, led by lead economist Ara Kharazian, publishes a monthly AI adoption reading drawn from the payments flowing through the product.

The index launched on 14 April 2025 as a section of Ramp's Spring 2025 Business Spending Report. Within 18 months it had become a reference series. Reporters quote it and a Fed database hosts it. It is also one of the four numbers compared in the comparison of four enterprise LLM market share trackers.

This post is not about whether the number is right. It is about the machine around the number: how the report is built, how it travels, and what it plausibly does for Ramp's demand.

My position up front: the index is the best B2B research marketing programme running today, and most companies that try to copy it will fail. The reasons it works are specific.

How the Ramp AI Index research report is built

The methodology is published in a short post titled How we built the Ramp AI Index, by Kharazian and Ian Macomber, Ramp's head of analytics engineering and data science. It is dated the same day as the launch. That timing matters, and I will come back to it.

The adopter definition is one sentence long

A business counts as an adopter if it has a transaction for an AI product or service in a given month. Ramp identifies those transactions from merchant names and line-item text on receipts and bills. At launch, the post described extracting that text from billions of anonymised transactions across more than 30,000 businesses.

The definition is deliberately narrow. It measures payment, not usage, satisfaction or spend. That narrowness is what makes it reproducible every month without a survey, and it is also what makes it easy for a reporter to explain in one clause.

The method also positions itself against the incumbent measure. The launch post argues that surveys can lead to underreporting and calls the Census question broad and abstract. Picking a fight with the official number is a choice, and it is a good one, because it gives every release a built-in comparison.

The sample grew and the limitations were stated on day one

The sample has grown with Ramp's customer base. The launch post cited more than 30,000 businesses. Press coverage in May 2026 cited more than 50,000. Ramp's June 2026 update and the FRED series notes now cite more than 70,000 US businesses.

The launch post named its own weaknesses. It said critics will point out that Ramp customers are more likely to adopt AI anyway, and that the data misses free tools and personal AI accounts. Stating the sampling bias up front costs nothing and pre-empts the first hostile question from any journalist.

Very few vendor reports do this. Most bury the sample in a footnote or omit it. A methodology page published alongside the first number is the cheapest credibility asset a research report can have. The survey-design side is covered in the analysis of the methodology gap in AI ROI surveys.

The distribution stack behind the Ramp AI Index

A number is not a channel. Distribution is. Ramp built four layers of distribution around the index, and only one of them is owned media.

Subscriber count and Bloomberg Terminal availability are Ramp's own statements, from its January 2026 lab update. FRED hosting was verified directly on 18 October 2026.

FRED hosting is worth more than any press hit

FRED is the St. Louis Fed's economic data service. The Ramp series sit there under IDs like RAMPAIALL, with a release labelled "Ramp AI Index (Not a Press Release)" and a line reading "Citation required."

That listing does three things a press placement cannot. It puts the number next to official statistics. It makes the data downloadable and chartable by anyone, so every analyst chart becomes a Ramp citation. And it compounds, because a FRED series gets cited for years after the article that introduced it is forgotten.

I would trade ten feature stories for one FRED listing. Most B2B companies never ask whether their data could qualify.

A FRED listing is not a Fed endorsement. The Fed's own April 2026 note on AI adoption, by Jeffrey S. Allen, relied on the Census survey and other sources and did not cite Ramp in the portion I could read. Hosting and endorsement are different things, and the distinction matters when you describe your own reach.

A recurring release beats a launch

Most research marketing is a single report, a launch week and a long decline. The Ramp AI Index is a series. Each month produces a new reading, and every reading can move.

A series turns a research report into a beat that reporters return to. Kharazian's January 2026 post set out plans for more datasets on macro trends and fast-growing companies, which extends the same logic to more beats.

How the press picked up the Ramp AI Index on 13 May 2026

The clearest single case is the May 2026 release. On 13 May, Kharazian published "Anthropic beats OpenAI on business adoption" on the Ramp Economics Lab newsletter. It reported that Anthropic reached 34.4% of businesses in April against OpenAI's 32.3%, the first time Anthropic had led.

The same day, The Decoder ran the figures with a link to the index. VentureBeat published "Anthropic finally beat OpenAI in business AI adoption" that afternoon. Axios ran "Anthropic overtakes OpenAI in workplace AI adoption." Let's Data Science ran the story and said Kharazian had told TechCrunch he was sceptical the lead would last.

Why the headline wrote itself

Three ingredients made the story portable. A named contest between two companies readers already follow. A clean first-time event. And a quotable economist, whose line that he had never seen a software industry as dynamic appeared in both The Decoder and VentureBeat.

Note what the coverage did not need. No press release, no visible embargo list, no paid placement. The release went out on a newsletter and the press did the rest, because the news value was in the data.

Every outlet repeated Ramp's caveat. VentureBeat noted the index counts card and invoice payments and likely underestimates adoption. The Decoder noted it tracks the share of companies paying, not how much they use. Publishing your weaknesses first means the press repeats your framing of them, rather than discovering them.

This is the opposite of the vendor survey playbook, where a striking statistic is engineered for the headline. That pattern and its credibility costs are covered in the breakdown of the Edelman B2B thought leadership study.

What the Ramp AI Index does for demand generation

Here the evidence thins, so I will separate what is documented from what is inferred. Ramp has not published any figure linking the index to pipeline, signups or brand search volume.

The product line follows the data

On 26 June 2026, Ramp relaunched the index. The June update said it had shifted from simple business adoption to tracking intensity of adoption, adding spend per employee and the split between subscriptions, coding agents and tokens. It reported a median of $11.38 per employee per month and $7,449 at the top 1% of firms.

About three weeks later Ramp launched AI Token Spend Management. CFOtech reported on 20 July 2026 that Ramp built it with more than 1,300 businesses managing more than 100 trillion tokens a month. It found potential savings of 12% of monthly AI spend for the average user.

The sequence is the demand engine in plain view. The research report tells every finance leader that AI spend per employee is rising and uneven. The product sells control over exactly that spend.

That is the version of demand generation worth copying, as distinct from lead capture, a distinction laid out in the comparison of demand generation and lead generation. Nobody fills in a form to read the Ramp AI Index. It is ungated, and the people it persuades arrive later through channels the index cannot claim credit for.

Brand search: what can and cannot be measured

What does the index do for Ramp's brand search? Nobody outside Ramp can measure it. Google Trends gives relative interest with no attribution. Paid keyword tools estimate volume, not cause.

What is measurable from outside is citation. Every press article in the May cycle linked to ramp.com/data. Every FRED download carries the Ramp name. Those are referring links and mentions, and they are the inputs that the AI search and attribution debates both depend on, as argued in the piece on dark discovery and attribution.

Index effects would be hard to isolate anyway. Ramp announced a $750 million round at a $44 billion valuation in June 2026. It reported more than $1 billion in annualised revenue and more than 70,000 customers, per FinTech Global's report of the announcement. A research programme is a small input into a company that size.

The research report as a channel: a working model

Strip out the fintech specifics and the Ramp AI Index reduces to five design choices. Each one separates a research report that becomes a channel from one that becomes a PDF nobody opens twice.

The last row is a real concession. Transaction data shows behaviour with no explanation. A survey can ask a buyer why they switched, and the Ramp AI Index cannot.

The gating row deserves emphasis. A gated report trades reach for a contact list, and for a data series that is a bad trade, because the reach is the asset. The arithmetic of that trade is worked through in the economics of gated content forms.

The voice row is the one most companies get wrong. Kharazian said publicly he was sceptical Anthropic's lead would persist, in the same news cycle that announced it. A marketing team would have cut that line. Keeping it is why reporters trust the next release.

Where this case is weakest

A case study that only praises is an advertisement. Here is where the Ramp AI Index, and this argument for it, falls short.

The famous headline did not survive revision

The May 2026 story said Anthropic passed OpenAI in April. In June, Ramp said it had made methodological updates to better capture enterprise spend on OpenAI and Anthropic. The current FRED series now shows April 2026 at 38.2% for Anthropic and 39.4% for OpenAI. In the revised data, OpenAI still led in April. Anthropic first leads in May, at 40.7% against 39.3%.

The direction of the story held, since Anthropic did overtake within a month. The specific claim that made the headlines, a first lead in April, is not what the series says today. I found no correction appended to the coverage I read.

This is the real risk of a monthly series. Every revision is a chance for a critic to say your headline was wrong. Ramp's June post disclosed the method change, which is better than most would do. It did not spell out that the crossover month had moved.

The sample is a customer list, and you are probably not Ramp

Ramp customers are businesses that chose a modern spend platform. Ramp said on day one that they are likely more AI-inclined than average. A reader comment on the May post by Josh Bersin argued the sample may skew toward tech or small firms and asked for demographics to be disclosed. The visible thread did not answer him.

The bigger limit is transferability. The index works because Ramp sits on a payment stream touching tens of thousands of businesses, monthly, with merchant-level detail. Most B2B companies have nothing comparable. A product analytics vendor with 400 customers cannot publish a national adoption rate, and pretending otherwise produces exactly the thin vendor statistics this site keeps flagging.

I also cannot show that the index drives revenue. The product sequence is suggestive. The press pickup is documented. The link from either to Ramp's pipeline is an inference, and anyone claiming a number for it from outside the company is guessing.

How to build a proprietary data content programme without card data

You do not need payment rails. You need a measurement your product already takes, that nobody else can take, about a question people outside your customer base care about.

Start by listing the events your product logs that describe your customers' behaviour rather than your own. Support tickets resolved, invoices sent, jobs posted, deploys shipped, contracts signed. Then ask which of those would interest a reporter who has never heard of you. Most will not. You are looking for one.

Pick a single metric with a one-sentence definition, the way Ramp picked "paid an AI vendor this month." Publish the methodology page before the first number, with the sample, the time window and the exclusions. Name the bias you know about. Then commit to a cadence you can sustain for two years, because a series that stops after three releases tells the press you were running a campaign.

Finally, publish the data in a downloadable form. A CSV on your own site is the minimum. The broader case for why first-party numbers beat restated ones, and what a survey-based version costs, is made in the argument that original research is the only content moat left. The Ramp lesson is narrower: the moat is the series, not the report.

If your honest answer is that you have no such metric, do not manufacture one. A commissioned survey dressed as an index will be read as marketing, because it is. Better to publish nothing than a number you would not defend under a reporter's second question.

Frequently asked questions

Where to start this week

Open your product's event log and write down every customer behaviour it records, one per line. Next to each, write the one-sentence definition a reporter could repeat. Cross out every line that only your customers would care about.

If one line survives, draft its methodology page before you compute a single number. If none survive, you have saved yourself an expensive campaign, and that is a useful finding too.

Related reading