AI Insights

Not just answers. Next steps.

Every answer comes with a plain-English summary of what changed, why it matters, and what to try next — grounded in your actual data, never fabricated.

AI summary

Paid social CPA dropped 18% this week, driven mostly by a creative refresh on the "Spring launch" LinkedIn campaign. Google Ads CPA was flat. Overall blended CPA is tracking at $42.80, down from $51.20 last week.

Recommended next steps

  • → Shift $2k of budget from Google Display to LinkedIn for the next 7 days.
  • → Test 2 more variants of the winning ad copy.
  • → Review landing page bounce rate — up 6% over the same period.

Three layers of insight, every answer

Plain-English summary

A concise explanation of what the data shows, written against the actual query result.

Next-step recommendations

Concrete actions you can take, informed by the numbers in front of you — not generic advice.

Anomaly detection

When something moves meaningfully, Meaning flags it and explains the context.

Examples from real questions

Paid channel comparison

Google Ads ROAS was 4.2×, LinkedIn was 2.1×. Google Ads delivered twice the return per dollar this month.

Recommended next steps

  • Shift 15% of LinkedIn budget to Google Ads high-intent keywords.
  • A/B test two new LinkedIn creatives before scaling back up.

Funnel drop-off

Mobile checkout drop-off at step 2 jumped from 24% to 38% last week — right after the latest deploy.

Recommended next steps

  • Roll back the form validation change from last Tuesday.
  • Add a funnel alert so drop-off > 30% pages you next time.

SEO opportunity

You have 14 queries with >5,000 monthly impressions but CTR below 2% — mostly listicles ranked positions 4–7.

Recommended next steps

  • Rewrite title tags for the top 5 underperforming pages.
  • Schedule a Search Console CTR alert for ongoing monitoring.

Email performance

Open rates on the last three campaigns dropped to 18.4% — previously running at 23%. Subject line length increased 40% in the same window.

Recommended next steps

  • Test shorter subject lines (< 45 characters) next send.
  • Pull best-performing historical subjects as templates.
Accuracy first

Meaning refuses to fabricate.

LLM-based analytics has an accuracy problem. Meaning solves it by passing the real query result to the model and giving it strict guardrails — if the data isn't there, the answer says so.

  • Summaries are generated from the actual query result
  • No invented numbers, ever
  • Explicit refusals when data isn't available
  • Every claim is traceable to a row in the result
Grounded in real rows
Reproducible query
Source-aware definitions
Traceable claims
No fabrication — ever

How the model stays honest

Query first

The SQL runs before the model writes anything — the model only ever sees real numbers.

Strict prompting

The model is instructed to refuse rather than invent, and to cite the row for every claim.

Delta-aware

Week-over-week, period-over-period comparisons are computed, not guessed.

Questions

Are the recommendations grounded in real data?+

Yes. Every insight is generated directly against query results — no synthetic numbers, no hallucinations. Meaning is built to refuse rather than invent.

Can I turn insights off?+

You can ignore them — answers always include the chart and table too. The AI summary is additive, not a replacement for the raw data.

Does the AI remember context across messages?+

Yes. Insights build on the previous message, so follow-up summaries stay consistent with the thread.

Which model generates the insights?+

Claude, chosen for its accuracy and safety characteristics — both critical when numbers can't be fabricated.

Does Meaning detect anomalies automatically?+

Yes. When a metric deviates meaningfully from its recent trend, Meaning calls it out in the summary and suggests a next step.

Ready to try Meaning?

Start your 14-day free trial. Cancel anytime.