Prompting an AI Writer for Buyer Guides: What to Put in the Brief
Most "AI prompts for product reviews" are a paragraph of adjectives: engaging, authoritative, SEO-friendly, 2,000 words. The model reads that and produces exactly what it would have produced anyway.
What changes the output is a brief with decisions in it: which searcher, which products, which voice, which claims are off-limits. Six inputs cover it, and they're the same six whether you're typing into a chatbot or filling in a tool's form.
This post sets out the six inputs and why each one changes the result, a filled brief for a real keyword, the things people put in prompts that make guides worse, what's different when the tool supplies the product data, and why the brief matters more than which model runs it. The brand voice post covers the voice input in depth; this one is the whole brief.
The short answer
Give the model six things: the keyword and the searcher's intent, the reader in one line, the product count and any products that must be in or out, the voice in a paragraph, the claims it may not make, and the facts it should work from.
Keep each to a few sentences. Leave out adjectives, word counts, "SEO-friendly", and anything you'd contradict later. The result is a guide you edit for fifteen minutes rather than rewrite.
The six inputs
| Input | What it decides | Weak version | Strong version |
|---|---|---|---|
| Keyword and intent | Page type and what the verdict answers | "bike lights" | "best bike lights for commuting in traffic; searcher wants to be seen, not to see" |
| Reader | Which trade-offs matter, what's assumed | "cyclists" | "daily commuter, 20 to 40 minutes each way, city streets, charges at a desk" |
| Product count and constraints | What's on the page | "top 10" | "8 products; must include a sub-$30 pick and a rear light; no helmet-mounted; no products with under 200 reviews" |
| Voice | How it reads | "friendly and professional" | Sentence length, person, verdict bluntness, hedging, humour, each with an instruction |
| Banned claims | What it can't say | (nothing) | "no testing claims; no prices; no spec not in the data; no 'durable/versatile/sleek'" |
| Facts | What it writes from | (nothing; model remembers) | Listing data per product, owner-review themes, or the tool's live data |
Keyword and intent
The keyword alone tells the model the topic. The intent tells it what the verdict answers and which spec leads.
"Best bike lights for commuting in traffic, to be seen" produces a guide ranked on side visibility and daytime modes; "best bike lights for unlit roads" ranks on lumens and beam. Same products, different page. One sentence of intent is the highest-leverage line in the brief.
Reader
One line describing a person, not a segment. "Daily commuter who charges at a desk" tells the model that USB-C matters and a 20-hour runtime doesn't. "Cyclists" tells it nothing. The reader line is what makes the trade-off sentence in each product section specific.
Product count and constraints
How many, and the rules. A budget pick required, a category excluded, a minimum review count, a product that must be in because you own it, a product that must be out because it's discontinued. These are the decisions the product count post describes, given to the model up front rather than fixed afterwards.
Voice
A paragraph setting the five dimensions with instructions a model can act on: sentence length with a number, person, how bluntly the verdict is stated, which hedging words are banned, whether humour is allowed. Set once per site; reuse for every brief.
Banned claims
The shortest input and the one that prevents the most damage. No testing or hands-on claims unless you supply them. No prices. No spec that isn't in the facts supplied. No adjectives from the ban list. No "best overall" without a reason in the same sentence. A model told what it can't say leaves gaps rather than filling them, which is what you want.
Facts
With a general chatbot, this is the hard half: you paste the listing data for each product (title, variant, specs, box contents, owner-review themes), because without it the model writes from memory and the specs are wrong. With a tool that pulls live product data, this input is supplied for you, and the brief is the other five.
A filled brief
Keyword: best bike lights for commuting. Intent: the searcher rides in city traffic and wants to be seen; side visibility and daytime flash matter more than raw lumens. Reader: commutes 20 to 40 minutes each way on lit streets, charges at a desk, rides year-round.
Products: 8, including one rear light and one under $30; exclude helmet mounts and anything with fewer than 200 reviews; include the Cygolite Hotshot (we own it). Voice: [site voice paragraph].
Don't: claim testing beyond the Hotshot; print prices; state a runtime or lumen figure not in the data; use "durable", "versatile", "sleek", "perfect for". Facts: [listing data per product, or the tool's live data].
About 120 words, and every line changes something in the output.
What makes a prompt worse
- Word counts. "2,500 words" produces padding to hit the number. Set the product count and let the length follow.
- "SEO-optimised" and keyword density. Produces keyword stuffing. Give the keyword once in the intent line.
- Adjectives as instructions. "Engaging, authoritative, conversational." None of them tells the model what to do differently.
- A famous writer's style. You get a parody.
- Contradictions. "Warm and friendly" plus "no hedging, blunt verdicts". Pick one.
- Asking for opinions. "Include your personal experience." It has none; it will invent some. That's the editing line, not a prompt line.
Chatbot vs tool
In a chatbot, all six inputs are yours to type, every time, and the facts input is the work: finding and pasting listing data for eight products takes longer than the writing.
In a tool built for affiliate guides, the facts come from live product data, the voice is set once per site, the banned claims are built in, and the brief per guide shrinks to the keyword, intent, reader and product constraints, which is a minute. The output quality is set by the same six things either way; the tool just supplies the three hardest.
Doing it with AffSERP
AffSERP's X Best Review form is the six inputs as fields: keyword, product count, the product constraints (must-include, exclusions, minimum reviews, price band as a tier), and the per-site voice brief, with the banned claims and the specs-over-adjectives rule applied to every guide and the product facts pulled live from Amazon.
The intent and reader lines go in the brief field in a sentence each. The editing post covers what happens after: the fifteen minutes that add what no brief can.
FAQ
What should an AI prompt for a product review include?
The keyword and what the searcher wants, who the reader is, how many products and which must be in or out, the voice, and the claims it isn't allowed to make (testing, prices, invented specs). With a tool that pulls product data, that's the whole brief; with a general chatbot you also have to supply the product facts, which is the harder half.
Does a longer prompt produce a better buyer guide?
Up to a point, then it gets worse. A brief that sets the six inputs in a short paragraph each beats a two-page prompt full of adjectives and contradictions. The model follows concrete instructions ('verdict in the first sentence') and ignores vague ones ('make it engaging').
Should I tell the AI which product to rank first?
If you know, yes, and say why in a line; the guide reads as yours. If you don't, let it rank from the data and change it in the edit if you disagree. What you shouldn't do is leave the model to invent reasons for an order you imposed; give it the reason.
Does the model matter more than the prompt?
Less than the brief and far less than the data. The same brief on two current models produces guides that differ in texture, not in whether they're right; a bad brief on the best model produces a confident generic page. Fix the inputs first.
Keep reading
