Every brief we write at Choco Media starts with research — competitor analysis, audience pain points, keyword clusters, platform benchmarks. Before we built our ai research workflow around LLMs, that phase would easily eat two to three days. A strategist hunting through browser tabs, stitching together a picture from ten different sources, cross-checking numbers, writing a summary no one would read in full. Now we do the same quality of first-pass research in two to four hours. This post is about exactly how we do that — the prompting patterns, the source-checking layer, and where we still do things by hand.
This is not a pitch for any particular tool. We use Claude, ChatGPT, and Perplexity at different stages, and the choice matters less than the method. If you run briefs, pitch decks, audience reports, or competitive analyses as part of your marketing work, this is written for you. You will leave with a repeatable process you can adapt for your own team.
A note on scope: “research” here means discovery and synthesis — understanding a market, an audience, a competitor set, or a topic well enough to make a creative or strategic decision. It does not mean academic research or anything requiring primary data collection. LLMs are useful for the former and unreliable for the latter.
Why the old research process was slow — and what changed
The bottleneck in traditional research is not finding information. It is reading, pattern-matching, and synthesising across many sources into a usable format. A typical competitive analysis might require reading eight to twelve company websites, three or four industry reports, several Reddit threads, and a handful of review sites. Each source takes time. The synthesis at the end takes more.
LLMs change this in two ways. First, they can compress public knowledge into a starting framework in minutes — not a finished document, but a solid scaffold. Second, they are very good at reformatting raw material you paste in: turning a wall of text into a structured brief, extracting themes from interview transcripts, or comparing two positioning statements side by side.
What they cannot do reliably is tell you what is true right now. Prices change, companies pivot, market share shifts. Any LLM output that contains specific numbers or recent claims needs to be verified against primary sources. This sounds obvious, but it is the step most teams skip, and it is why AI research gets a bad reputation.
- LLMs compress synthesis time by 60–80% on discovery tasks
- They are strongest on framing, structuring, and comparing — not on current facts
- Source-checking is not optional; it is the last 20% that makes the output trustworthy
- The best use is as a first-pass researcher, not a final authority
The four research tasks we run through LLMs every week
Not all research is the same. We have found four recurring tasks where LLMs consistently save meaningful time without introducing unacceptable risk.
1. Audience pain-point mapping
We paste in raw material — Reddit threads, review site text, customer interview transcripts, support ticket summaries — and ask the model to extract recurring pain points, organise them by theme, and rank them by frequency. This used to take half a day of reading and sticky-noting. It now takes twenty minutes of pasting and prompting, plus thirty minutes of checking the output against the source material.
2. Competitor positioning analysis
We copy the key pages from three to five competitor websites — home, about, main service or product page — paste them into a prompt, and ask for a side-by-side comparison of positioning claims, tone, and differentiators. The model produces a structured table we can work from immediately. We still visit the sites ourselves to catch anything the copy misses (design language, trust signals, pricing structure), but the written positioning layer is done in under an hour.
3. Topic and keyword landscape
Before we brief any content, we use an LLM to map the semantic neighbourhood of a topic — what sub-topics exist, what questions audiences are likely asking, what angles have been covered to death versus what feels fresh. We then verify the most important keyword volumes using a real SEO tool. The LLM gives us the map; the tool tells us which roads are worth driving.
4. Brief generation from raw inputs
When a client sends us a brief that is actually a brain dump — a long email, a deck with forty slides, a voice note transcript — we run it through an LLM to extract the core objective, audience, constraints, and success metrics. The output is a one-page structured brief we can use to align the team without another meeting.
- Audience pain-point mapping from raw text
- Competitor positioning side-by-sides
- Topic and keyword landscape mapping
- Brief structuring from messy client inputs
The prompting patterns that make this reliable
There is a difference between prompts that produce plausible-sounding output and prompts that produce useful, verifiable output. In our ai research workflow, we have settled on a small set of patterns that consistently deliver the latter.
Role + task + format + constraint
Every research prompt we write follows this structure. Role: what the model should act as (“You are a senior marketing strategist”). Task: what we need it to do (“Analyse the following competitor positioning statements”). Format: how the output should be structured (“Produce a markdown table with columns for: brand name, core claim, tone, primary audience, and one gap or weakness”). Constraint: what it should avoid or flag (“Do not invent information not present in the text I provide; flag any claim you are uncertain about”).
The constraint line is the most important. Without it, models will fill gaps with plausible-sounding fabrications. With it, you get honest uncertainty flags you can then resolve with a quick check.
Paste-and-extract over ask-and-receive
We get far better results when we give the model the source material rather than asking it to recall information from training. “Here are five competitor About pages — extract their positioning” beats “Tell me how these five companies position themselves” every time. The former is synthesis; the latter is recall, which is where hallucination risk is highest.
Iterative layering
We rarely try to get everything in one prompt. We use a chain: first pass extracts raw themes, second pass organises them, third pass identifies implications or gaps. This mirrors how a good analyst would actually work, and it gives us checkpoints where we can correct the direction before we are deep into the wrong framework.
“The model is best used as a thinking partner that compresses reading time — not as a source of ground truth. The moment you treat it as the latter, the research becomes unreliable.”
The source-checking layer we never skip
Every piece of research that leaves our team passes through a brief verification step before it is used in a client deliverable or internal decision. This is not as slow as it sounds — the LLM output tells us exactly what to check. Any specific number, named tool, quoted stat, or factual claim gets a quick primary-source verification. We keep a shared doc where we log the source URL and date for each verified claim.
For competitor research, we always cross-reference the LLM’s extracted positioning against a live visit to the page. Positioning changes. What was true six months ago may not reflect where a competitor is now. For audience pain-point work, we check that the themes the model extracted actually appear in the source text — this catches both hallucination and over-generalisation.
Our rule of thumb: if a fact would embarrass us if it turned out to be wrong, we check it. If it is directional rather than specific (“this audience tends to care about speed and simplicity”), we treat it as a hypothesis and test it in the work rather than burning time verifying the obvious.
- All specific numbers verified against primary sources
- Competitor positioning checked against live pages
- Audience themes checked against source material the model was given
- Sources logged in a shared verification doc
Where we still do research by hand
Honest answer: anything that requires judgment about what is not being said. LLMs are good at surface pattern recognition. They are not good at the kind of lateral thinking that produces genuine strategic insight — noticing the positioning gap that exists because all the competitors are afraid of it, or recognising that an audience pain point is really a symptom of a deeper unmet need.
We also do primary research by hand. Client interviews, user surveys, expert calls — nothing replaces talking directly to people. LLMs can help us prepare better questions and synthesise the output, but the conversations themselves need a human in them.
Anything that requires reading between the lines of a brand — understanding the cultural subtext of how a company talks, or what a tone of voice choice signals about who they are trying to be — we still do collaboratively as a team. This is the kind of nuanced judgment that LLMs flatten into generalities.
Our AI content creation service sits downstream of this research process — the briefs we produce using this workflow feed directly into the content production layer. The research quality is what determines whether the content has anything real to say.
How to build this into your agency workflow
The practical setup is simpler than most people expect. You do not need a complex stack. What you need is a consistent process and a team that understands where the tool is reliable and where it is not.
Start with one task
Pick the research task your team does most often and runs least efficiently. For most content-producing teams, this is competitor analysis or audience research. Build a prompt template for that one task — role, task, format, constraint — and use it consistently for four weeks. Refine based on what you find the model gets wrong or misses.
Build a prompt library, not a prompt collection
A collection is prompts saved somewhere. A library is prompts that are organised, named, versioned, and annotated with when to use them and what to watch out for. We wrote about this in more depth in our post on building an AI prompt library your team will actually use — the short version is that the organisation layer is what makes a prompt library durable rather than a folder that goes stale.
Make source-checking a team norm, not an individual habit
The verification step only works if everyone does it, not just the careful people. We built it into our workflow as a required step before any research leaves a strategist’s hands — not a suggestion, not a “when you have time.” The log doc is visible to everyone, which creates light accountability without creating overhead.
Review outputs as a team occasionally
Once a month, we review two or three research outputs in a team session — not to fact-check them again, but to ask whether the LLM framing missed anything important that a human would have caught. This keeps us honest about the method’s limits and usually surfaces one or two improvements to our prompting approach. Our AI automation service is built on exactly this kind of iterative refinement — you do not optimise a workflow once; you tend it.
- Start with one recurring research task, not the whole process
- Build a prompt library with annotations, not just saved prompts
- Make source-checking a team norm with a visible shared log
- Monthly team reviews of LLM research outputs to catch method drift
What this looks like in practice: a real brief cycle
To make this concrete: here is roughly how a client strategy brief moves through our process. The client sends us a scope of work and a brain dump. We paste the brain dump into a structured prompt and extract the core brief in ten minutes. We identify the three to five competitors we need to understand, copy their key pages, and run a positioning extraction — thirty minutes. We run an audience pain-point extraction against two or three Reddit communities and a handful of reviews — another thirty minutes.
At that point we have a draft framework: what the market looks like, where the competitors are positioned, and what the audience cares about. A strategist then spends an hour with that framework — checking it against their own knowledge, making live visits to the sites, adding the lateral observations the model cannot make. Source-checking takes another thirty minutes.
The output is a research brief that used to take two full days and now takes roughly half a day. The quality is comparable or better, because the strategist is spending their time on judgment rather than reading.
The limits you should know before you rely on this
LLMs have a training cutoff, which means recent market changes, new product launches, pricing updates, and current events will be missing or wrong. For research in fast-moving categories — AI tools, advertising platforms, consumer trends — this matters a lot. Always treat LLM output as a starting point and assume the most time-sensitive elements are stale.
Models also have a bias toward coherent-sounding outputs. They will produce a well-structured competitive analysis even when the underlying data is thin or contradictory. The structural confidence of the output can mask uncertainty about the content. This is why the constraint line in your prompts matters: explicitly asking the model to flag uncertainty produces more honest output than letting it fill gaps silently.
Finally: this approach makes research faster, not infallible. Faster research that is poorly verified is worse than slow research that is right. The speed gain is valuable only if the source-checking layer is genuine.
- Training cutoff means recent facts will be stale — check everything time-sensitive
- Structural confidence can mask thin or uncertain underlying data
- Speed gain is only valuable with genuine source verification
- Models work best on analysis of provided material, not recall of external facts
If you want to talk through how this kind of workflow could fit your team, or you are looking at what a research-to-content pipeline might look like end to end, we are happy to get into the specifics. Reach out here and we can set up a short call.