92% blog with AI. Only 13.9% get results.

News: On September 8, 2026, Andy Crestodina published the new edition of Orbit Media's annual survey, based on 1,042 content marketers. Two numbers sit side by side in it: 92.4% of respondents now use AI to blog, and only 13.9% report strong results, the lowest reading in the survey's twelve-year history (Orbit Media, September 8, 2026).

For twelve years this survey has worked as the industry's thermometer. The share of marketers reporting strong results had always sat between 20% and 30%. It has just dropped to 13.9%, six points below the previous floor. Over the same period, production has never been faster. That simultaneity is what deserves a close look.

Four years of productivity gains, and a historic floor

The report documents the speed gain precisely. Average writing time per post is now 3 hours and 20 minutes, down from a peak of 4 hours and 10 minutes in 2022, with four consecutive years of decline. At a normal publishing cadence, that adds up to dozens of hours saved per person per year.

It is worth being strict about what this data proves and what it does not. A self-reported survey does not demonstrate that AI degrades performance: it measures two curves moving in parallel. Crestodina himself makes no accusation against the tool. Search Engine Land, covering the study on September 10, keeps the same nuance: what separates programs that work is not AI adoption, it is whether the practices that already worked have been dropped.

The most profitable practice is the one collapsing

This is the most useful finding in the report, and the least discussed. Among all the practices tested, one stands out with a multiplier this clear: marketers who regularly collaborate with experts or outside contributors beat the benchmark by 2.6 times on reported results.

And that practice is in retreat. It covered 25% of respondents in 2017. It covers 7% today. The lever most correlated with success has been cut by more than two thirds in nine years, while the most widely adopted tool reached 92% penetration.

One last number closes the argument. The report isolates eight practices associated with the strongest-performing programs, then counts how many respondents do all eight: zero out of 1,042. Nobody does the complete job. That is not a knowledge problem, it is a time allocation problem.

Why this matters even more in generative search

On Google, average content could still pick up long-tail traffic. Inside an answer generated by ChatGPT, Perplexity or AI Overviews, there is no long tail: there is a handful of cited sources, or nothing. The engine has no reason to cite a page that restates what it can already produce on its own.

What these systems go outside for is precisely what they cannot generate: a figure measured in a specific setting, a date, a methodology, a named author. Exactly what expert collaboration produces, and exactly what this survey shows disappearing. We documented the same mechanism in our analysis of the study of 220 sites publishing AI content at scale, and Google had already settled the underlying debate: it is not AI content that gets penalized, it is bad content.

What to do this week

1. Compute your own ratio. Across your last twelve months of publishing, count the posts that produced a ranking, a citation or a lead, and divide by the total published. Above 13.9% and you are beating the panel average. Below it, volume is not your problem.
2. Reallocate the time saved instead of converting it into volume. The fifty minutes saved per post should not fund one more post. They should fund the thirty-minute interview with the person who actually knows, the one who supplies the one sentence nobody else can write.
3. Put an expert back in the loop, even one hour a month. That is the report's 2.6 multiplier, and it is also the foundation of what generative engines reward when they pick sources. A salesperson, a technician, a long-standing customer all qualify: expertise does not need to be famous, it needs to be real and attributed.
4. Change the metric you steer by. Track the share of your strategic queries where your domain shows up as a cited source, not just how much you published. That is the starting point of the method we use to appear in ChatGPT and AI Overviews.

The limits of what we know

Three caveats before generalizing these numbers.

• This is a self-reported survey. "Strong results" are graded by respondents themselves, with no shared definition and no analytics verification. Part of the decline may reflect rising expectations as much as falling performance.
• The report establishes correlations, not causation. The 2.6 multiplier attached to expert collaboration does not rule out that already high-performing teams are simply the ones who can afford to bring experts in.
• The 1,042-respondent sample is predominantly English-speaking and the report publishes no country-level breakdown. None of these percentages can be presented as a figure for any specific national market.

The Cicero take

The dominant story of the last two years was that AI would industrialize content. This report shows it industrialized writing, which is not the same thing. Writing was never the bottleneck. Knowing something others do not, having measured it, being able to sign your name to it, that is the bottleneck, and it does not automate. The 13.9% is not the price of AI, it is the interest on a debt taken out by removing the human step the machine does not replace. The right way to use these tools is to make them pay for that step, not delete it: software-grade productivity in service of agency-quality work, not instead of it.

Frequently asked questions

Alexis Dollé

CEO & Founder

Growth and SEO & GEO content strategist, I founded Cicero to help businesses build lasting organic visibility, on Google and in AI-generated answers alike. Every piece of content we produce is designed to convert, not just to exist.

LinkedIn