The most dangerous moment in any report is when a number moves and everybody assumes they know why. Three stories from the past week show the scoreboard changing for reasons that have nothing to do with the work.
Each one is small on its own. Together they make a case for asking how a number is produced before celebrating or panicking about it.
Your PageSpeed score went up and you did nothing
According to Search Engine Journal, a Lighthouse update has changed how PageSpeed Insights simulates page loads. Google's Barry Pollard said the new version improves the simulated First Contentful Paint and Largest Contentful Paint, which should be more accurate and generally produce higher scores.
The practical consequence is in one sentence from the piece. A page may receive a different PageSpeed Insights score without anything on it changing. JavaScript-heavy sites are the most likely to see a jump.
Lighthouse 13.6 changed parts of Lantern, the simulation behind those estimates. It now uses Chrome's own render-blocking label for each request, and it estimates text-based LCP differently when the largest element appears in the same frame as the first content.
I can already picture the agency report. "Performance improved 18 points this month." Nothing was shipped. The ruler changed length.
This is the difference between lab data and field data. Lighthouse runs a simulated load on one device with fixed network conditions. The Chrome User Experience Report measures what real users actually experienced. If your lab score jumped and your field data did not, your customers did not notice anything.
The update also changed how Lighthouse looks for Agentic Resource Discovery manifests, now checking for ard.json and an Agentmap line in robots.txt. That is a small detail, and a useful reminder that the audits themselves are being rewritten for AI agents.
AI answers are getting harder to read
The second signal comes from Lance Eliot in Forbes, covering a study of the psycholinguistic features of AI output. It found that frontier models now use more complex structures than human writers, including more low-frequency word combinations and stacked nouns.
One plausible explanation is that models are increasingly optimised for goals like maths and coding rather than for communicating clearly. The result could be a readability gap, where average users struggle to understand or act on what the AI tells them.
Here is why this belongs in a piece about measurement. Dense, technical prose looks more intelligent. A manager skimming an AI-written report sees long words and confident structure and scores it highly, without checking whether the reader could actually do anything with it.
For marketing teams, the risk is direct. AI-drafted copy can drift towards writing that impresses the person approving it and loses the customer reading it. Eliot's suggested fix is simple: instruct the model to write at a specific reading level without sacrificing accuracy. I gave a fuller version of that advice in write like an aircraft manual.
Half the agent insights still go nowhere
The third signal is a report from MIT Technology Review Insights on agentic analytics. On average, 48% of insights generated by analytics agents at the surveyed companies lead to actions that improve the business. A year ago, respondents put that figure at 37%.
That is progress. It also means more than half of the insights are still not helping anyone. A small group the report calls insight leaders gets more than 70% of agent insights to useful action.
The causes are not about the models. Lack of business context was cited by 55% of executives as the main reason for poor insights, and fragmented or inaccessible data by 45%.
The overload number is the one I would put in front of a board. At 52% of organisations, agents generate too many insights to evaluate properly. Only 24% can evaluate all of them. The agent is producing output faster than the company can read it, which turns the insight count into another vanity metric.
The companies improving fastest are working on governance, cited by 60%, and adding a semantic layer to their data architecture, cited by 45%. In plain terms, they are teaching the agent what the business means by revenue, customer and churn before asking it for conclusions.
The common thread
All three stories describe a gap between the metric and the reality it is supposed to represent. A score that rose because the test changed. Writing that looks smarter while communicating less. Insight counts that grow while action stays flat.
I covered a related pattern in your metrics broke, not your content. The response is the same each time. Before you report a number, write down how it is produced, and check it against a second number that measures what the customer experienced.
A dashboard is a model of the business. When the model changes, the business does not change with it.