PR teams improve LLM analysis output by giving the model a written brief, supplying the source text, asking for evidence next to every claim, running the analysis in separate steps and checking a sample against the original. A language model produces fluent text whether or not the text matches the source, so quality comes from the instructions and the review loop, not from the model's tone.
What does this article cover?
Why does LLM analysis output go wrong in PR work?
LLM analysis output goes wrong because a language model generates the most likely next words given its instructions and input, and it does so whether or not the result matches the source. A coverage summary can read well and still misstate a quote, merge two outlets or invent a trend.
Four causes account for most PR errors. The question is vague, so the model chooses its own scope. The source text is missing, so the model fills gaps from general patterns. The output has no required structure, so claims arrive without evidence. The task is too large, so details drop out between the start of the answer and the end.
Each cause has a fix inside the prompt or the workflow. The wider role of the person who owns that quality bar is described in the AI Workflow Editor role.
How do you write a brief an LLM can follow?
A good brief states the task, the audience, the source material, the output format and the limits in plain sentences before any analysis starts. The model has no knowledge of the client, the reporter or the deadline unless the brief supplies it.
A workable brief for a coverage analysis covers six items:
- The job, for example: summarize how the five attached articles describe the client's product recall.
- The reader, for example: a communications director who has not seen the articles.
- The source text, pasted or attached in full, with outlet names and dates.
- The output format, such as a table with outlet, date, main claim and tone.
- The limits, such as: use only the attached text and write "not stated" when the text does not say.
- The definition of terms the team uses, such as what counts as positive, neutral or negative coverage.
The "use only the attached text" instruction matters most. It tells the model to work from the supplied articles instead of general patterns, and it gives a reviewer a clear test: every claim must appear in the source.
How do you make an LLM show evidence for every claim?
An LLM shows evidence when the prompt requires a short supporting excerpt and its location next to every claim in the output. A claim with no excerpt is a claim the reviewer must verify from scratch.
The prompt can ask for a table where each row carries the claim, the outlet, the date and the exact sentence from the article that supports it. Keep quoted excerpts short, since the goal is to point the reviewer to the right place in the source and not to reproduce articles. If the model cannot supply an excerpt, the prompt should tell it to mark the row "no supporting text found" instead of guessing.
Two further instructions help. Ask the model to separate what the sources state from what it infers, so inference is visible. Ask it to flag conflicts, for example when two outlets report different figures, instead of choosing one. Both instructions turn hidden errors into visible ones, and visible errors are cheap to fix.
Why should you split analysis into separate steps?
Splitting analysis into separate steps keeps each model run small enough that details stay accurate and each result can be checked before the next step builds on it. A single prompt that asks for extraction, comparison, sentiment and recommendations at once gives errors room to hide.
A three-step sequence works for most coverage and research tasks:
- Step one extracts facts: outlet, date, reporter, claims and quotes, into a table with no interpretation.
- Step two analyzes the table: themes, differences between outlets and gaps against the client's key messages.
- Step three writes the summary for the reader from the checked table, in the team's usual format.
The team checks the table after step one, because every later step inherits its mistakes. Splitting also lets the team assign different model tiers to different steps, a cost lever covered in cheaper AI models for PR tasks.
How should a PR team review LLM analysis output?
A PR team should review LLM output by checking a sample of claims against the original sources, logging each error by type and revising the brief when the same error repeats. Review that stops at "it reads well" misses the failures that matter.
A practical review loop has four parts. Pick a sample of claims, such as every number plus a handful of others. Trace each to the source article. Record each error as a wrong figure, a misattributed quote, an unsupported inference or an omission. Then change the brief to prevent that error type and rerun a few examples. The log also shows whether a task is ready for a smaller, cheaper model or still needs a top-tier one.
Prompt libraries speed up the first draft of a brief. ChatGPT prompts for PR teams pairs each prompt with a check to run, and Perplexity prompts for media research covers research queries where sources are shown. The full workflow around these steps appears in the AI PR stack.
Which failures weaken PR analysis, and how do you fix each?
The table below lists five common failures in LLM analysis for PR, with the likely cause and the change that addresses it. Treat it as a starting checklist and add the errors your own log reveals.
| Failure | Likely cause | Fix |
|---|
| Quote attributed to the wrong outlet | Sources pasted without labels | Label each source with outlet and date, and require a quote plus outlet in the output |
| Trend that no article states | Open-ended request for insights | Require every claim to cite a supporting excerpt and separate inference from fact |
| Wrong or rounded figure | Numbers summarized from memory of the text | Ask for figures in a table with the exact source sentence, then check each one |
| Generic summary that fits any client | No client context in the brief | Add the client's key messages and ask for gaps against them |
| Detail lost from long inputs | One prompt handling too much text | Split into extraction, analysis and summary steps |
For the deck stage that often follows an analysis, see building a client PowerPoint with AI, which treats the checked table as the deck's source of truth. Teams comparing tools for these steps can start from the AI tools for communications teams guide.
