Essay
The Gap Was the Format
Why “nobody reads our research” is almost never a reading problem, and what to change before you commission the next report.
· 6 min read
Somewhere in your organization there is a report. It cost real money. Someone interviewed forty customers, or benchmarked fifty competitors, or modeled three years of category data. It runs eighty pages, it is correct, and almost nobody has opened it.
When that happens, marketing teams reach for one of two explanations. The first is that the audience is lazy. The second, more common and more expensive, is that the report didn’t get enough reach, so the answer is to produce more of them: a blog series, a webinar, a carousel, another report next quarter.
Both explanations skip what the evidence points at. The bottleneck was the container the research arrived in.
The experiment this industry should know and doesn’t
In 1991, Bruce Britton and Sami Gülgöz published a study in the Journal of Educational Psychology that deserves to be better known outside education research. They took a single 1,000-word expository text and used a model of reading comprehension to find the places where the text quietly asked the reader to supply a missing connection. Then they repaired those places, supplying the connections the text had left for the reader to make.
Then they tested it twice. With 170 undergraduates, free recall of the repaired version was much higher than recall of the original. With 125 Air Force recruits, they looked at the structure readers took away.
The part that matters for a marketer is how they decided what “the author’s intentions” were. They didn’t ask a focus group what it preferred. The original author and seven subject-matter experts described the structure the text was meant to leave in a reader’s head. Their structure correlated above .5 with that of recruits who read the repaired version, and only .1 with those who read the original. The authors’ own conclusion is hedged: the revision “apparently” carried out the author’s intentions better than the original text. By that measure, what readers took away from the repaired version looked much more like what the author and the experts intended.
What changed was how the text connected its ideas.
Now take away their time
That study wasn’t about time pressure. Your readers are almost always under it.
In 2017, Feifei Han published a study in Frontiers in Psychology with 402 second-year undergraduates reading in a language that wasn’t their first. Each of them read under both conditions: once with forty minutes on the clock, once with no limit. Comprehension scored 12.99 under time pressure against 15.14 without it, a difference the paper reports as F(1,401) = 127.48, p < .01. One thing to know before leaning on that: the two conditions used different passages, and the paper doesn’t report counterbalancing, so part of the gap could come from the texts rather than the clock.
There’s a second finding in that paper that’s more useful than the first. Under time pressure, how efficiently a reader recognized words correlated with whether they understood the text, r = −0.22. With no limit, that correlation was no longer significant, r = −0.09, p = 0.07. Given enough time, weaker readers seem to compensate. Take the time away and they can’t.
Every professional audience you write for lives in the forty-minute condition permanently. They read between meetings, on a phone, with two other documents open. Hand that reader a format built for someone at a desk with an hour, and you remove the one thing that was quietly covering for the format.
And they may never arrive at all
Until recently you could at least assume that someone who wanted your report would find it and land on your page. Three measurements, using three different methods, say that assumption is weakening.
Pew Research Center studied it directly. Athena Chapekis and Anna Lieb analyzed 68,879 Google searches made by 900 US adults in March 2025, using a panel that records the pages people visit rather than asking them to remember. When an AI-generated summary appeared in the results, users clicked a traditional search result on 8% of visits. When no summary appeared, 15%. They clicked a source cited inside the summary on 1% of visits. And they ended their browsing session entirely on 26% of pages carrying a summary, against 16% of pages without one.
The Reuters Institute for the Study of Journalism found the same shape by asking. In its Digital News Report 2026, 4% of all respondents across 27 markets, a figure that includes people who don’t use chatbots at all, said they always or often click through to the original source from an AI chatbot. From search engines the figure was 19%, and from social media 17%. The three numbers share a denominator, which is what makes the comparison worth something. It’s also self-reported behavior, a weaker instrument than a browsing panel. Two methods with different weaknesses pointing the same direction is worth more than either on its own.
For context rather than proof: SparkToro, the software company run by Rand Fishkin, published an analysis in June 2026 of a Similarweb browsing panel covering January to April 2026 in the United States. It found 68.01% of Google searches ended without any click, against 60.45% in 2024. That’s industry research from a company with a commercial interest in the subject, not an institute, and Fishkin notes his own number is probably an undercount because the panel misses Google’s mobile app. I’d treat it as a reading on the direction of travel and not lean on it.
Put those together and a long report looks like a bet on three things at once: that the reader would arrive, that they’d stay, and that they’d read in the order you wrote. All three have become weaker bets, and producing more reports doubles down on all three.
Someone else is writing your summary
There’s a practical consequence in the Pew numbers that’s easy to read past. On pages carrying an AI summary, 1% of visits included a click on a source the summary cited, and 8% a click on an ordinary result. Most visits included neither. Pew can’t say what those readers took from the page, only that they didn’t click through. The closest it gets to ‘stopped there’ is the session figure: 26% of pages with a summary ended the browsing session, against 16% without.
Marketing teams have spent two decades getting good at controlling the first thing a reader sees. The headline, the subject line, the opening frame. More and more, that first thing is a summary someone else’s system wrote, working from your eighty pages without asking which part mattered.
None of the research here measures how those systems treat different kinds of documents, so what follows is reasoning, not a finding. A document that saves its conclusion for the last page gives any summary less to work with than one that states its claim up front and then supports it. The second kind is also the kind a reader in a hurry can use.
It’s close to what Britton and Gülgöz found from the other direction: how a text is built shapes what leaves the page with the reader. What’s changed is that the reader is no longer the only one doing the reading.
Why the wrong answer keeps winning
There’s a reason “make more” survives contact with evidence. A report is easy to commission, easy to count, and easy to show a board. Restructuring one is none of those things: it looks like less work and costs more judgment. Nobody gets credit for the eighty pages that didn’t get written.
So the format usually gets chosen by default, which means it gets chosen by whoever is most comfortable with long documents. That’s usually the person who already knows the material, and they’re probably the least able to feel the joins, because they supply the missing connections without noticing.
What the evidence adds up to
Strip the three findings down and they point the same way. Repair the places where a reader has to guess at a connection, because a text with fewer gaps to fill was remembered better. Design for a reader who is short of time, because that’s the condition in which weaker readers stop being able to compensate. And put the claim where a hurried reader, or a summary, will meet it first, because when a summary appears, fewer people click through to the source.
All of it comes down to building the document for the moment it will be read. None of it requires writing more.
Five questions before you commission the next one
Five questions worth putting in a brief, or asking before anyone approves something long.
What moment is this read in, and how many minutes does that moment have? If nobody can answer, the format is being picked by default.
Does the first screen say what we found and what it means, or only what the document is about?
If a reader only ever sees a summary of this, written by someone who isn’t you, what has to be in the first forty words?
What are we producing instead of the eighty pages, not in addition to them? “In addition” is how a format problem becomes a budget problem.
Who here knows the material best, and why is their read on clarity the one to question first?
None of this is about making things shorter for its own sake. The eighty pages can exist, and sometimes they have to: the credibility of a short version depends on the long one being real and available. The mistake is shipping the long version as if it were the product, then reading the silence that follows as a lack of interest.
“Nobody reads it” is almost never a reading problem. More often, the format assumed a reader who was going to arrive, sit down, and start at the top.
Sources
Peer-reviewed research
Britton, B. K., & Gülgöz, S. (1991). “Using Kintsch’s Computational Model to Improve Instructional Text: Effects of Repairing Inference Calls on Recall and Cognitive Structures.” Journal of Educational Psychology, 83(3), 329–345. A 1,000-word expository text revised at the points where it required reader inference. With 170 undergraduates, free recall of the revision was much increased over the original. The original author and 7 subject-matter experts provided the intended cognitive structure; their structure shapes correlated above .5 with those of US Air Force recruits who read the revision and only .1 with recruits who read the original. The authors conclude that the revision “apparently” carried out the author’s intentions better. The ERIC record gives 125 recruits and describes two experiments. doi.org
Han, F. (2017). “Applicability of the Compensatory Encoding Model in Foreign Language Reading: An Investigation with Chinese College English Language Learners.” Frontiers in Psychology, 8, 681. Study 1: N = 402 second-year undergraduates, within-subject design, so every participant read under both conditions. Comprehension M = 12.99 (SD 2.88) with a 40-minute limit against M = 15.14 (SD 2.42) unrestricted, F(1,401) = 127.48, p < .01. Word recognition correlated with comprehension under time pressure (r = −0.22, p < .01) and not without it (r = −0.09, p = 0.07). The two conditions used different passages and counterbalancing is not reported. Open access. doi.org
Research institutes
Chapekis, A., & Lieb, A. (2025). “Google users are less likely to click on links when an AI summary appears in the results.” Pew Research Center, July 22, 2025. 68,879 unique Google searches by 900 US adults, collected March 1–31, 2025; 18% of the searches in the study produced an AI summary. Measured with a panel whose members installed a tracker that records their browsing, so this is observed behavior, not self-report. pewresearch.org
Ross Arguedas, A. (2026). “Emerging uses of AI chatbots for news and what it means for journalism.” Digital News Report 2026, Reuters Institute for the Study of Journalism. Fielded by YouGov. The click-through figures cover 27 markets and, in the report’s words, are “for the full sample, meaning they include non-users”, so the 4% is a share of all respondents, not of chatbot users. Self-reported behavior. reutersinstitute.politics.ox.ac.uk
Industry research, labeled as such
Fishkin, R. (2026). “In 2026, Less than One Third of Google Searches Still Send a Click.” SparkToro, June 9, 2026. Similarweb desktop and mobile web panel, United States, January–April 2026. A commercial publisher writing about its own market; the author notes the figure likely understates the trend because the panel excludes Google’s mobile search app. Used as context, not as support for the argument. sparktoro.com
