51% of Marketers Use AI Agents. Only 6.5% Measure If They Work.

Here is a stat worth sitting with. In the last year, the share of marketing organisations using AI agents went from one in five to roughly one in two. In the same survey, the share that can say whether those agents are actually working sits at one in fifteen.

That is not a typo, and it is not a small sample fluke dressed up as a trend. It is the headline finding of a report that has been tracking this industry for four years running, and this year's edition landed on 19 October 2026.

Adoption outran measurement. By a lot. Here is what the data actually says, where the gap is widest, and what a marketing team without a data science function can realistically do about it.

The numbers, in order

Start with adoption. 98.4% of respondents now use generative or agentic AI in some form. That number has been near-universal for a while and is no longer the interesting one.

The interesting one is 51.1% of organisations have created or started using AI agents — tools that act on a task rather than just answering a prompt. A year earlier, that figure was 20.0%. It did not grow. It more than doubled.

Daily AI use climbed too, from 33.6% of respondents in 2025 to 61.4% this year. Premium tool spending is now the norm, not the exception: 88.6% of organisations pay for at least one paid AI tool.

Then the measurement section. Asked how their organisation tracks AI's return on investment, respondents split five ways:

• 6.5% measure it against defined metrics and benchmarks — roughly 12 of the 184 people surveyed.
• 23.4% measure it only anecdotally.
• 18.5% do not measure it yet, but plan to.
• 22.7% do not measure it at all.
• 23.9% are not sure whether anyone at their organisation measures it.

Add those up. Nearly two-thirds of respondents are either not measuring AI's return, or do not know if their organisation is. That is the gap the headline number describes.

Why the agent number matters more than the usage number

A chatbot answering one prompt at a time is easy to reason about. You wrote the prompt, you read the output, you decide if it was useful.

An agent is different. It is set up once and then acts repeatedly — drafting a week of social posts, triaging inbound leads, adjusting a media budget — without a person reviewing every step. That is exactly the kind of system where a silent failure can run for weeks before anyone notices.

Basis's own subgroup here is small: agent-use questions drew 94 of the 184 respondents, and the leadership-specific questions draw on 63 to 65. Treat the precise decimal points as directional, not exact. The shape of the finding — adoption surging, measurement barely moving — is what carries across a survey this size, not the fourth digit.

This is not a one-survey story

If the Basis numbers stood alone, they would be worth a raised eyebrow and not much more. They do not stand alone.

A separate study of 333 marketing decision-makers, run by Datalily through Centiment for the email platform Knak and reported by MarTech in July 2026, found marketers were 68% more likely to measure email and landing-page performance by click-through rate than by revenue or pipeline influenced. 69% track clicks. Only 41% track revenue.

Different survey, different vendor, different 2026 fieldwork window. Same shape: it is easier to watch an AI system produce something than to prove that what it produced moved a number that pays the bills.

BuzzRiding's own 28 days of Search Console data is a small, local version of the identical problem. Impressions are a vanity-adjacent metric that is easy to watch rise. Clicks, and the conversions behind them, are the harder number, and the one that was stuck near zero.

The quieter number in the same report: Claude passed ChatGPT

Buried in the same Basis dataset is a shift that has nothing to do with agents and everything to do with which tool marketers reach for first.

ChatGPT usage among respondents fell from 88.6% in 2025 to 74.5% this year. Claude usage rose from 17.9% to 72.3% — a four-fold increase that puts it within two points of ChatGPT. Microsoft Copilot sits at 55.4%, Gemini at 47.8%.

Salesforce's own Agentforce, notably, was used by just 1.6% of respondents, down from 5.0% a year earlier — a reminder that a vendor's AI agent brand and a marketer's actual agent usage are not the same chart. BuzzRiding's own comparison of ChatGPT, Claude and Gemini for marketing tasks is worth revisiting in light of that swing; a tool ranking from July 2026 was written before Claude's usage share had quadrupled.

Why so few teams measure agent ROI

None of the sources cited here explain the mechanism in survey-question form, so this section is inference from the shape of the data, not a reported cause. Three plausible, non-exclusive reasons line up with what both surveys show.

First, agents are new enough that most teams are still at the "does it work" stage, not the "how well does it work, measured against what" stage. 51% adoption a year after 20% adoption means a large share of agent users are, by definition, in their first year.

Second, the metrics that are easy to pull — content produced, tasks completed, hours apparently saved — are not the metrics that prove value. The Knak/MarTech data shows marketers already default to the easy metric (clicks) over the hard one (revenue) even for channels that have existed for a decade. An 18-month-old category is not going to buck that habit faster.

Third, defined benchmarks require a baseline from before the agent existed, and a lot of agent deployments are net-new workflows with no "before" to compare against. You cannot measure the lift from automating a task nobody was doing systematically last year.

What this means if you are the one deciding whether to keep paying for the agent

You do not need a data science team to close this gap. You need three things most marketers already have: a number from before, a number from after, and a date you agreed to check both.

1. Pick one agent, one workflow, one metric — before you expand it. Not "AI is helping with social." Pick the specific agent, the specific workflow it touches, and the one number that workflow is supposed to move.
2. Write down the baseline before the agent starts, not after. If a lead-triage agent goes live on the 1st, your baseline is response time and conversion rate for the 30 days before the 1st — recorded on the 1st, not reconstructed from memory in week six.
3. Set the comparison metric to something revenue-adjacent, not activity-adjacent. "Posts published" and "emails drafted" are activity. "Leads that became opportunities" and "click-through rate on CTA, not just open rate" are closer to revenue. The Knak/MarTech data above is the cautionary case for defaulting to the easy number.
4. Put a date on the recheck and keep it even if the news is bad. 22.7% of Basis respondents do not measure AI ROI at all, and 23.9% do not know if anyone does. Both of those are default outcomes of skipping this step, not active decisions.

This is, essentially, the same four-step discipline BuzzRiding applies to its own Search Console numbers every month: state the metric, record the baseline, publish the result whether it flatters you or not, repeat on a fixed schedule.

The obvious caveat

These are two surveys, not a census. Basis drew on 184 respondents, with agent-specific questions answered by 94 of them; Knak's study drew on 333 respondents at companies with at least $50 million in revenue, a different population than most small marketing teams. Neither publisher disclosed a margin of error. Percentages built on subgroups under 100 people move several points on a handful of answers either way. The direction — adoption climbing faster than measurement — is consistent across two unrelated studies; the exact decimals are not load-bearing.