--- title: "Most Marketing Insights Are Just Summaries: How to Write One That Is Not" description: "A summary arranges what happened. An insight names a cause, commits to an action and says how confident you are. Here is the test, and the reporting habit that produces them." author: "Team" category: "Marketing Insights" date: 2026-10-18T14:35:00.375Z canonical: "https://contentagents.dev/blog/most-marketing-insights-are-just-summaries" --- # Most Marketing Insights Are Just Summaries: How to Write One That Is Not ![telescope, insight, outlook](https://cdn.pixabay.com/photo/2013/06/08/00/21/telescope-122960_1280.jpg?w=1200&q=75) > A summary arranges what happened. An insight names a cause, commits to an action and says how confident you are. Here is the test, and the reporting habit that produces them. A chart tells you what happened. A deck that only arranges charts has done the describing and left the thinking to whoever reads it last. That is why so many monthly reports land well, get praised in the meeting, and change nothing by the following Monday. The test proposed here is narrow and unforgiving. Can you finish this sentence: "because of this, we are going to do that"? If the sentence will not close, you have written a summary. Summaries have their uses. They are simply not what a leadership team meant when it asked for insight. ## Reporting and learning are two different jobs Reporting answers the question "what happened". Learning answers "what should change". The two draw on the same data and almost none of the same skills. A reporting cycle rewards completeness, so it grows every quarter: more channels, more tabs, more slides. A learning cycle rewards one good question, so it shrinks. Teams default to the first because tooling made description nearly free while leaving interpretation exactly as expensive as it was. Research from the Content Marketing Institute puts the gap plainly: around nine in ten B2B marketers now use AI somewhere in content creation, while the share reporting that it actually improved performance sits well under half of them ([Content Marketing Institute](https://contentmarketinginstitute.com/press-room/new-b2b-research-finds-winning-marketing-teams-are-building-fundamentals-not-just-playing-with-prompts)). Output is easy to make and easy to count. Judgement is neither. ## What a well-made finding looks like Borrow the shape from research that shows its working. An analysis published by Temso looked at how AI systems choose sources by language, covering roughly 350,000 AI responses and 7,058,891 individual citations gathered across 12 countries and 47 industries in early 2026 ([Temso](https://www.temso.ai/data/-Lost-in-Translation-How-AI-Models-Handle-Local-Language-Sources)). It names the systems, the prompt languages and the window. That transparency is precisely what makes its headline arguable. The headline: citations matched the language of the prompt in 85.4% of cases for Google AI Overviews, 70.2% for ChatGPT and 51.7% for Grok ([Temso](https://www.temso.ai/data/-Lost-in-Translation-How-AI-Models-Handle-Local-Language-Sources)). Notice what is absent. There is no claim about why the systems differ, and no claim that the pattern will hold next quarter. The authors measured, bounded and stopped. Copy the format rather than the topic. A finding worth acting on carries four things: the population it covers, the window it covers, the comparison that gives the number meaning, and a limit the author volunteers before anyone demands it. Remove any one of the four and what is left is a statistic looking for a slide. ## The data pipe breaks before the thinking does Before blaming the analysis, check whether the record still exists. A standard Google Analytics property retains event-level data for either 2 or 14 months. The longer windows of 26, 38 and 50 months belong to 360 properties only, and the very largest properties are held at 2 months ([Google Analytics Help](https://support.google.com/analytics/answer/7667196)). That setting is usually chosen once, by whoever connected the property, and it quietly decides whether a year-on-year comparison is available to you at all. Search Console carries its own ceiling. Its Search Analytics API returns a maximum of 25,000 rows per request, and the reference states outright that it does not guarantee every row, only the top ones ([Google Search Console API reference](https://developers.google.com/webmaster-tools/v1/searchanalytics/query)). So the long tail of phrasings your buyers really type is partly missing from the export you are reasoning over. Our walkthrough of [content intelligence with GA4 and Search Console](/beyond-the-pageview-engineering-content-intelligence-with-ga4-gsc-8abd) covers how to pull it without pretending it is the whole picture. Two consequences follow, and both are jobs for this week. Raise the retention setting now, because it does not reach backwards into history. And write the sample beside every figure in the report, so a reader can see how much of the world you are actually holding. ## Two published findings, read properly ### Clicks: first the number, then the decision Pew Research Center followed the real browsing behaviour of 900 adults in the United States through March 2025. On results pages carrying an AI summary, a traditional result link was clicked during 8% of visits; on pages without one, 15%. Links inside the summary itself were clicked in roughly 1% of visits, and sessions ended on 26% of pages carrying a summary against 16% of those without ([Pew Research Center](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/)). That is four figures, one country, one month, one search engine. The decision hiding inside them: if your highest-intent queries now carry a summary, a falling click line is partly a change in the surface rather than a failure of your content, so the report needs an impressions line and a branded-search line beside it before anybody draws a conclusion. That sentence is the insight. The four figures are its evidence, not its replacement. ### Correlation: the strongest signal is still not a lever Ahrefs examined 75,000 brands and ranked how tightly each signal tracked visibility in Google's AI Overviews. Branded web mentions led at 0.664, branded anchors followed at 0.527, branded search volume at 0.392, and backlinks came last at 0.218. Louise Linehan, who wrote the study up, holds the line that correlation is not causation ([Ahrefs](https://ahrefs.com/blog/ai-overview-brand-correlation/)). Compressed into a summary, that becomes "mentions drive AI visibility", which the study does not support. Turned into an insight, it becomes a bet of a stated size: brands people discuss tend to show up, so spend one quarter deliberately earning mentions and watch whether your own curve bends. The first version is an assertion. The second is a test you are allowed to lose. ## The same month, told twice Below is one month at a fictional B2B site, described two ways. No figures appear in either version, because what matters here is the shape of the sentence rather than the size of the move. - **As a summary:** "Organic clicks were down month on month. Impressions were up. Our strongest query is still content operations software. Newsletter opens held steady." - **As an insight:** "Clicks fell while impressions rose, and the queries behind that split are the ones that now carry an AI summary. Two sales calls this month mentioned hearing our name from an assistant before reaching the site. So we are adding a branded-search line to the weekly view and shipping the comparison page buyers keep asking for. Confidence: moderate. Evidence: a dozen queries and two calls." The second version can be wrong, and that is the entire value of it. It names a mechanism, commits to an action, and admits how thin its evidence is. Nobody argues with a bar chart. People argue with a claim, and the argument is where a team learns anything. ## The three-finding limit is the part worth defending Most advice on this subject tells you to get better at interpretation. Our position is narrower and less comfortable: cap the report at three findings and let good material go unreported. The cap is not a preference for tidy documents. It is the only mechanism that reliably forces ranking, and ranking is where interpretation actually happens. A team permitted twenty findings will produce twenty descriptions, because describing is quicker than deciding and nothing in the format punishes the substitution. The research above points the same way. Those studies hold up because each commits to a small set of claims and shows how it got there. The vendor figures that circulate without a sample fail for the opposite reason: adding one more cost the author nothing. A three-item limit imports that cost into your own cycle. Something has to be cut, so somebody has to argue about what matters, so the argument finally gets had in a room instead of being deferred to the reader. The cap earns its keep except in two situations. If a month's findings are genuinely independent, as they can be in a business running several unrelated product lines, then three per line beats three in total. And if your team already reviews last month's decisions and finds the actions were taken and the numbers watched, the cap is fixing a problem you do not have. Keep it while it hurts, because the discomfort is the mechanism working. On choosing what to count in the first place, see our note on [the ROI metric B2B teams get wrong](/the-content-roi-metric-most-b2b-teams-get-wrong-vi4d). ## The routine that produces them - **Choose the question before opening the dashboard.** One question per month. The dashboard answers it, and it will never generate it for you. - **Close the sentence.** "Because of this, we will do that." No closed sentence, no finding. - **Attach one human source.** A line from a call, a ticket, an interview. Numbers show you where to look; a person tells you why. - **Declare confidence and sample.** Strong, moderate or weak, plus what it rests on. - **Audit last month before adding this month.** Did the action happen? Did the metric move? Skip this and you are collecting trivia with excellent formatting. Competitive work answers to the same discipline. Mapping what a rival publishes earns its time only if it ends in a choice, which is the framing in our guide to [reverse-engineering a competitor's content strategy](/how-to-reverse-engineer-a-competitors-content-strategy-in-90-minutes-nyms). For the measures that belong at each stage, see [B2B content marketing metrics by funnel stage](/b2b-content-marketing-metrics-by-funnel-stage). ## If the sentence will not close, the report is not ready Everything before "because" is data. Everything after "we will" is a commitment. The insight is the join between the two, with a confidence level attached so a reader knows how hard to push. Hold your own reporting to that join, and most of what currently fills your deck turns out to be an appendix you were presenting as an argument. ## FAQs about turning marketing data into insights ### What exactly counts as a marketing insight? A statement that explains a behaviour and commits to a change. Data is the measurement. A summary arranges measurements. An insight names a likely cause, proposes a next action, and declares how confident you are in both. ### Where do the strongest ones come from? Usually from pairing a pattern in the numbers with a reason from a person: a sales call, a support ticket, a customer interview. The dashboard finds the anomaly. The conversation explains it. ### How many belong in a monthly report? Three at most, with everything else moved to an appendix. A report carrying twenty findings hands the prioritisation back to its reader, which was the work the report existed to do. ### Can an AI tool produce them for you? It can surface patterns and draft the description quickly. Judging whether a pattern is real, material and worth acting on stays with a person who carries the consequences. Treat any generated finding as a candidate, never as a conclusion. ## FAQ ### What exactly counts as a marketing insight? A statement that explains a behaviour and commits to a change. Data is the measurement. A summary arranges measurements. An insight names a likely cause, proposes a next action, and declares how confident you are in both. ### Where do the strongest ones come from? Usually from pairing a pattern in the numbers with a reason from a person: a sales call, a support ticket, a customer interview. The dashboard finds the anomaly. The conversation explains it. ### How many belong in a monthly report? Three at most, with everything else moved to an appendix. A report carrying twenty findings hands the prioritisation back to its reader, which was the work the report existed to do. ### Can an AI tool produce them for you? It can surface patterns and draft the description quickly. Judging whether a pattern is real, material and worth acting on stays with a person who carries the consequences. Treat any generated finding as a candidate, never as a conclusion. --- Source: https://contentagents.dev/blog/most-marketing-insights-are-just-summaries