Build a prompt plan

This guide is for brands and agencies setting up prompt tracking in GEOflux. It covers what to track, how much, in whose words, and how to keep it honest month after month.

Steps 1 to 5 happen once before anything is measured; step 6 repeats every month, and only the quarterly review reopens steps 4 and 5.

Why prompts decide everything you measure

Every number in GEOflux, Visibility, Share of Voice, Mentions, Citations and Sentiment, is computed against one thing: your prompt set, the questions GEOflux asks ChatGPT, Gemini and Perplexity on your behalf, every day on Starter and Growth. The prompt set is not a setting you configure once. It is the working definition of what "being visible in AI" means for your brand.

That has three consequences. A weak set produces confident, wrong numbers: track questions nobody asks and you measure a conversation that is not happening. Changing the set mid-quarter resets your baseline, so every month-on-month comparison breaks. And for agencies, the set is the scope of the engagement: a client who wants 50 more prompts is asking for more engagement, and the plan they are on is what sizes it.

The order that works: define the questions first, agree them with whoever owns the result (your team or your client), and only then start tracking. Do it in that order and every later conversation is about performance, not about what the number means.

A good prompt set mirrors how real buyers ask AI assistants, in their words rather than your marketing language. It covers the commercially important topics first and weights them to where the money is. It is mostly non-branded, because the fight is the category question ("best X for Y"), not the question where the buyer already typed your name. It is stable enough to trend and governed enough that every change is a dated decision. And it fits your credit budget with room to spare.

---

The basics

A prompt is not a keyword. A keyword is the 2 to 5 words someone types into Google ("personal loan rates"). A prompt is the full question the same person types into an AI assistant ("Which bank has the lowest interest rate on a personal loan right now?"). Real prompts are full sentences with context, not keyword fragments, and they get longer when the person is not using web search. Your prompts should read the same way; GEOflux accepts prompts of 4 words up to 200 characters, roughly 30 to 35 words.

How GEOflux meters your set

For the full credit model, see Billing and credits. For personas, see Personas.

The vocabulary of a prompt set

• Buyer Journey. Discover: the buyer is learning or describing a problem, with no options in mind yet ("how do I make good espresso at home", "my espresso tastes sour, is it the beans or the machine?"). Evaluate: the buyer is judging which options are worth considering ("best X for Y", "recommend me an X under 500", "X vs Y", "alternatives to X"). Buy: the buyer has decided what they want and asks where, whether or from whom ("where can I buy X online with free returns", "is X worth it", "which provider near me"). Rule of thumb: no options in mind = Discover; choosing between options = Evaluate; ready to act = Buy. "Where do I buy X" is a Buy prompt even though it returns a list of retailers. Answers that name brands happen mostly in Evaluate and Buy, so that is where most of the budget goes.
• Branded vs Non-branded. A branded prompt names you or a competitor ("is Monzo safe", "Monzo vs Starling"). A non-branded prompt names only the category. Branded prompts almost always mention the brand, so they flatter and teach little. Keep them to roughly 10% of the set, as an accuracy and sentiment check, and mark every prompt with the Brand type field so the split is visible.
• Prompt type. Category head ("best X"), use case or constraint ("best X for small teams under 100 dollars a month"), comparison ("X vs Y"), alternatives ("alternatives to X"), problem-led ("my espresso tastes sour, is it the machine or the beans"), local ("X near Manchester"), pricing and trust ("is X worth it").
• Set vs universe. The universe is every question your category asks (thousands). The set is the sample you track. Sampling is fine, because visibility is read as a share across prompts and days, never as a single prompt's position.

The two fields and how to read them

Give every prompt a Buyer Journey value (Discover, Evaluate or Buy) and a Brand type value (Branded or Non-branded). They are dedicated fields, not tags, so they do not use up tag slots, and the prompts table tells you how many prompts are still missing either one. General tags stay for the rest: product line, the month a prompt was added (added-2026-11) and the baseline marker (baseline-2026-q4).

How to read them in GEOflux: the headline number is the Overview with Brand type set to Non-branded, last 30 days, all engines. The funnel read is the dedicated Buyer Journey tab in the Overview. Filters combine with AND across fields and OR within one field, so the useful combined slices are a click away: Evaluate + Non-branded shows how you do when buyers compare options without naming anyone; Buy + Non-branded shows the purchase moment; Branded on its own is the brand-check slice. Filters only include, they never exclude, which is why every prompt needs a Brand type value: Non-branded is read by selecting it, not by removing Branded. See Filters and periods.

---

Step 1: Start from real demand

Nothing downstream can be more right than this foundation. Before writing a single prompt, collect five things.

See Track competitors for how to manage the competitor set.

Use Google data to size the conversation, then translate it

No AI engine publishes prompt search volume, and any "prompt volume" you see in a tool is a modelled estimate. Google keyword volume is the most reliable proxy you have: the same demand that types "cheap car insurance" into Google asks an assistant "what is the cheapest car insurance for a 25-year-old in the UK right now and where do I buy it". Volumes will not map one to one, but the ranking of topics by demand carries over.

The method we recommend for every brand (volumes below are illustrative UK figures):

1. Pull the head keyword of each topic with its monthly Google volume, for example "personal loan" (135,000), "car insurance" (450,000), "cheap flights" (1,200,000).
2. Under each head, list the variations that carry intent: "best personal loan rates" (14,800), "car insurance monthly payments" (2,900), "cheap flights to barcelona" (33,100). Modifiers such as best, cheap, online, price, reviews, vs, alternatives, for [segment] are the ones that turn into recommendation prompts.
3. Write one prompt per variation, in the buyer's voice, keeping the intent and the constraint. The keyword is the receipt for the prompt: it proves someone asks this.
4. Allocate prompts per topic roughly in proportion to head volume and commercial weight, not equally.

Other places real questions live

• Google Search Console: your top 50 to 100 non-branded queries, turned into questions. Longer, more specific queries are the ones most likely to already be answered by an AI feature.
• People Also Ask, Reddit, forums and review sites: the raw buyer voice ("is a 300 pound espresso machine actually better than a 150 one for a beginner").
• Sales calls and support tickets, plus a "how did you hear about us" field that asks for the exact prompt.
• GEOflux's AI prompt and topic suggestions. They are generated from your website and never mention your brand; they also take your tracked competitors, personas, existing prompts and topics into account, and are written in the language of your project's country. Treat them as a draft to judge, never as the final set. See Suggested prompts.
• The AI answers themselves: when an assistant answers a broad question it silently splits it into sub-questions about reviews, price, specs and comparisons. Each facet you see in a real answer is a prompt candidate.

---

Step 2: Build the topic tree and set the two ratios

Prompts are derived, not brainstormed. Your own site already holds the structure: the homepage states what you are (home coffee machines), the navigation lists the product families (espresso machines, bean-to-cup, pod machines, grinders, accessories), and each family splits into the situations people buy for (first machine, small kitchen, office, upgrade from pods). Each branch becomes a GEOflux topic, and each topic spreads across the three Buyer Journey stages. Add one more topic for comparisons and alternatives, which is where the branded prompts live.

The two ratios

These are starting defaults, not laws. The direction is solid: AI engines name brands far more often on comparison and buying questions than on educational ones, and in B2B categories many conceptual questions come back with no brand at all. How commercial a category's AI conversations are also differs a lot between industries, so tilt the split to your vertical and write the tilt down:

Inside each stage, keep a mix of prompt forms so that stage comparisons are not distorted by phrasing: roughly 40% list or "best" prompts, 30% constrained use-case or persona prompts, 20% "X vs Y" or alternatives, 10% open-ended. How a question is phrased can move brand mentions as much as which stage it belongs to.

On branded: give every prompt a Brand type. The headline Visibility and Share of Voice are read with Brand type set to Non-branded; that is how the brand-check slice stays out of the headline number. Filters include, they do not exclude, so the Non-branded value has to be on every non-branded prompt. Keep about 10% of the set branded, collected in a dedicated "Comparisons and alternatives" topic rather than scattered through the product topics, so the ratio holds at set level. On Growth, make it a deliberate brand-check slice of 10 to 15 prompts (accuracy, sentiment, "is X trustworthy", head-to-head against each direct competitor). Five branded prompts are too few for a sentiment trend; a marked slice gives you the sample without inflating the score.

Prompt types to cover inside each topic

A healthy product topic of 6 to 10 prompts mixes types rather than repeating "best X" with synonyms. Rewording a prompt without changing its intent barely changes the answer, so wording twins waste credits. Comparison and alternatives prompts, which name brands, go in their own topic (see the Starter table in Step 4), so the product topics stay non-branded.

And in the dedicated comparisons topic:

Weight the topics unequally

Allocation follows commercial weight. Weight to what makes you money (racing shoes get a fat slice, running socks a sliver). Weight to what you can deliver (no visibility on a line you cannot restock; that only converts into disappointed customers). And don't only fight the giants: a challenger tracking only head prompts is invisible, while the niche prompts are where depth beats scale. Coverage before depth: every commercially important topic gets Discover, Evaluate and Buy coverage before any topic gets a second layer.

Why topics matter for sentiment

Each prompt's topic is how GEOflux splits sentiment and brand perception by product line. A bank is rarely described the same way across everything it sells: AI may praise Larkmoor Bank's personal loans and call its credit cards expensive, while Pellinor Bank is the card it recommends. Pooled into one number the two banks look alike; split into a Loans topic and a Credit cards topic, Larkmoor leads on loans and Pellinor on credit cards.

With at least 5 prompts per topic, Competitor breakdown by topic on the Overview's Topics tab, set to Sentiment, shows which brand is best perceived on each topic among you and your tracked competitors. Brand Perception, filtered to one topic, shows why. So give each product line its own topic rather than one broad topic for everything.

See how topics split sentiment by brand in the video: Sentiment by topic.

---

Step 3: Write prompts the way people type them

AI engines answer the questions people ask, in people's language. A prompt written in your marketing vocabulary tests a conversation nobody is having. The one test that matters: would a real person type this into ChatGPT, exactly like this?

Ten rules

1. Write a question or a request, aim for 8 to 15 words, one intent per prompt. "Which bank has the lowest rate on a personal loan right now?", not "personal loan interest rate". GEOflux needs at least 4 words and accepts up to 200 characters, roughly 30 to 35 words.
2. Use the buyer's words, not the brand's. "Best coffee machine for a beginner", never "precision-engineered home barista system".
3. Add the context a real person adds: budget, location, use case, experience level. Constraints push the engine from explaining into recommending, and many real commercial conversations include one ("for a small team", "under 500 dollars").
4. Keep brand names out of non-branded prompts, yours and your competitors'. Mark the few branded prompts you keep as Branded.
5. Write in the language the market searches in, in a project set to that country. A German audience gets German prompts in a DE project; a UK audience gets British English in a GB project. Add a second language only if your audience really searches in it.
6. First person and natural imperfection are fine ("I want to buy...", "Recommend me..."). Many real prompts contain pronouns; keyword-shaped prompts under-represent how people ask.
7. One prompt, one job. Don't stack two questions; the engine answers one and you won't know which.
8. Skip pure definitions ("what is term life insurance") unless the slot is a Discover prompt you chose on purpose. Definitions rarely name brands.
9. Ask list- or ranking-shaped questions when you want brands surfaced ("what are the 5 best..."). Lists tend to surface more brands than open-ended phrasing. Don't make every prompt a list, though; real users don't.
10. Don't burn credits on wording twins. "Best CRM for small business" and "top CRM for SMBs" measure the same thing. Spend the slot on a different intent, constraint or persona.

Before and after

Personas: when they earn a credit

A persona is built from structured attributes (age, role, company size and so on) plus a short "Other details" note, and GEOflux adds it to the question as "about me" context, running it as a separate answer alongside the prompt's own answer. Changing the persona (role, company size, budget, family situation) can change a large share of the brands named, especially for smaller brands. That makes personas valuable and expensive: each link holds as many credits as a prompt (3 on the base engines). Link them to the 10 to 15 prompts where the audience genuinely changes the answer, not to the whole set. On Growth, 3 personas on 15 prompts is 45 links, or 135 credits, almost a third of the plan; 3 personas on every prompt turn 450 credits into about 37 prompts.

---

Step 4: Size the plan: 45, 150 and 450 credits

Credits are allocated, not consumed. A credit is one answer, so a prompt holds 3 credits (one each for ChatGPT, Gemini and Perplexity), and 3 more per linked persona, for as long as it is active; archiving returns them, restoring charges them again, and at each renewal the balance resets to the plan allowance minus what active slots hold. So the credit total is the size of your active set, and the plan is the size of the set. GEOflux generates reports from a single prompt, but the read only becomes meaningful from about 25 prompts. Below that, treat the set as a pilot of one topic, not a benchmark.

The shapes below are counted in prompts and slots, which is how you plan; multiply slots by 3 for credits. Every shape leaves 5 to 10 slots (15 to 30 credits) free. That headroom is what lets you add a prompt in month two without archiving one.

Free (45 credits, 15 prompts): a pilot

One topic, Evaluate and Buy prompts only, one market, 12 to 15 prompts. Free runs every 3 days and has no personas, which is one more reason it is a pilot, not a benchmark. It proves the mechanism (you see real answers naming your competitors) and seeds the paid set later.

Starter (150 credits, 50 prompts, 1 persona slot): one market, core topics

Forty-five prompts across 6 topics in one country is a sound monitoring set, with 5 slots (15 credits) free. Starter includes one persona, but every prompt you link it to takes a slot out of that headroom, so keep it to 3 prompts at most, or skip it.

Check: 10 + 10 + 7 + 8 + 5 + 5 = 45 prompts, 135 of 150 credits, 15 free.

What Starter cannot do well: more than one persona, and more than 6 topics. Those are the reasons to move to Growth. A second country or a second brand is never a reason to squeeze the plan: it is a separate project with its own plan.

Growth (450 credits, 150 prompts, 3 persona slots): pick one of three shapes

All three shapes are one brand in one country. A brand active in two markets runs two projects, each sized on its own plan, with the intent translated rather than the prompts copied word for word.

In every shape keep the 20/40/40 and 90/10 ratios at the level of the whole set, and keep the 5 to 10 free slots unallocated for the first quarter, so an addition never forces you to remove something.

Quick sizing formula

Base prompts = topics x 3 stages x 2 to 4 prompts per stage, then trim to the shape. Six topics x 3 x 2.5 = 45, the Starter set. Ten topics x 3 x 4 = 120, trimmed to 115, the prompts in Growth shape B. Twelve topics x 3 x 4 = 144, trimmed to 140, Growth shape A. If the formula gives more than the plan allows, cut topics before you cut stages: a topic with only Evaluate prompts is still readable, a topic with 2 prompts is not.

---

Step 5: Test, approve, baseline

Approve the set before it runs. In a company that means the marketing lead and the sales lead agree the list; in an agency it means the client signs it off. Either way it fixes the definition of visibility, stops moving goalposts (changes become dated decisions instead of a drip of "can we also track...") and, for agencies, draws the scope line.

Checklist before you import

• Every prompt has a demand receipt: a Google keyword with volume, a Search Console query, a forum thread or a support ticket behind it.
• Ratios hold at set level: about 20/40/40 Discover / Evaluate / Buy, about 90/10 non-branded, at least 5 prompts per topic.
• No wording twins. Two prompts that would get the same answer become one prompt plus a different intent or constraint.
• Language matches the market, and the project is set to the right country.
• 3 to 5 prompts pasted manually into ChatGPT, Gemini and Perplexity in an incognito session. If none of them names any brand, the prompt is in explaining mode; rewrite it to ask for a recommendation.
• Competitors named in those manual answers are in your tracked competitor set, within your plan's limit.
• CSV built for import: columns prompt, topic, buyer journey, brand type, tag 1 ... tag 5. Values are case-insensitive ("non branded" works); conflicts or unknown values are flagged in the preview. Topics and tags are created if new. Personas cannot be set in the CSV; assign them afterwards in the prompts table. The country comes from the project. Max 500 rows per file. GEOflux refuses the whole import if you do not have enough credits, and skips prompts you already track. Before you confirm, check the preview's "missing Buyer Journey" and "missing Brand type" counts; both should be zero. See Importing prompts from CSV.
• Credit total at or under the plan, with 5 to 10 slots (15 to 30 credits) of headroom.

The approval document

One page: the topic tree, prompt counts and weighting per topic, the two ratios, the competitor set, market and language, any deliberate deviation from the defaults with its reason, and the prompt list. Keep the CSV you imported as the canonical copy of the approved set: the prompt export in GEOflux contains ID, prompt, metrics, tags in one cell, persona, Buyer Journey, Brand type and the date added, but no topic column, so it is not in import format. When the page is agreed, you have a fixed definition of visibility.

Why day 2 differs from day 1

AI answers are not deterministic. Ask the same question twice on the same day and you will rarely get the identical list of brands; over a month, even a category leader is absent on some days, while smaller brands come and go. Rephrasing, engine updates, the user's location and language all move the answer. Nothing is wrong with your set when day 2 differs from day 1. That is the medium you are measuring.

On Starter and Growth, GEOflux runs every prompt daily on all three engines so that stability comes from aggregating across prompts and days rather than from any single run. Read visibility as a share across the set and across a window, never as a per-prompt position. See When data updates.

See Shared reports for how to create one.

Read every screen through the same filter

Every number is measured over a scope. The Overview filters are period (presets or a custom range), brand, engine, topic, tag, Buyer Journey, Branded vs Non-branded and persona. Within one filter, values combine with OR; across filters, with AND. The same visibility score can be excellent on one scope and poor on another. Before reading anything, set the period, the engine, the topic and the Brand type, then trust the number. When a score surprises you, change one filter at a time and watch which one moves it. For the funnel, use the Buyer Journey tab; for a slice, combine fields (Evaluate + Non-branded, Buy + Non-branded).

The non-branded read is the discovery read and usually the truer test: it shows how you perform when the buyer did not name you. Fix one headline scope (last 30 days, all engines, Brand type Non-branded) and use it every month, so numbers stay comparable and nobody asks "what is this measuring" twice.

---

Step 6: Keep it alive: the monthly and quarterly routine

A prompt set is a living thing, but a governed one. Refresh too often and you introduce noise and reset the baseline; never refresh and you track last year's conversation.

Weekly (15 minutes)

• Scoreboard read with filters set first: Visibility, Share of Voice, Sentiment and the competitor leaderboard, by topic and by engine. You are reading momentum, not reacting to a day.
• Sentiment: a negative movement is handled the same day, not at month end. On the Answers page, set a custom range of one day; each answer's drawer lists its cited sources, which link to the source pages in the app. Find the source, decide the response. The Answers page filters are period, brand, engine, prompt, sentiment, topic, tag, Buyer Journey, Branded vs Non-branded, persona and a text search. Set an alert so you don't have to watch for it: alerts fire on Visibility, Share of Voice or Sentiment, on a percentage change or a threshold, and arrive by email.
• New competitor showing up in GEOflux's suggestions with meaningful mentions: add it as a dated decision, within your plan's competitor limit.

Monthly (the close)

1. Flag what is not earning its credit. A prompt with zero brand mentions, zero citations and zero competitor appearances for 30 or more days is a question the engines answer without brands. Note it as a candidate for the quarterly review; do not archive it yet, because an archived prompt leaves the stats and changes the baseline read mid-quarter.
2. Keep what teaches. A prompt where competitors appear and you do not is the most valuable prompt in the set: it is a gap you can work on. Never prune those.
3. Add deliberately. New product lines, seasonal demand (Black Friday, summer holidays, back to school) and gaps found in the cited sources go in as dated additions, tagged with the month (added-2026-11), so you can read them apart from the baseline set. This is what the headroom is for.
4. Report at topic level, not prompt level. Per-prompt numbers move daily; 30-day Visibility and Share of Voice by topic are what you can act on.

A monthly report that keeps everyone aligned has four parts: what moved (a shared report with the month's custom date range, read against the baseline report, plus the Non-branded headline taken from the Overview with the same range, since Brand type cannot be shared yet), what we did (actions taken, with before and after), what we found (new risks and openings from the answers and sources), and what we need decided (the batched sign-off items). Run the meeting off that report: decisions first, then numbers, then next month's focus.

Quarterly (the set review)

• Re-check the ratios; sets drift toward branded and Evaluate prompts over a year of additions.
• Go through the candidates flagged at the monthly closes and the prompts flat for 90 days: archive, replace with a different intent, or investigate why nobody wins them. Archiving returns the credits; the history stays but leaves the stats.
• Re-pull Google volumes for the topic heads. Demand shifts; a topic that lost half its volume should lose prompts to one that gained.
• Re-run the manual incognito check on 5 prompts per engine. Engines change how many brands they name and how much they expand a question, so a prompt that surfaced brands in January may not in July.
• Present all changes as one dated amendment to the approval document. Changes mid-quarter reset the baseline; changes at the quarter boundary start a new chapter.

Rules of governance

• Additions and removals are dated decisions, because every change changes the test.
• Archive deliberately, and only at the quarterly review. An archived prompt leaves your stats and costs its credits again if you restore it. "Delete permanently" in the Archived tab erases the prompt and all its answers and cannot be undone; never use it on anything that was part of a baseline.
• Tag the original set (baseline-2026-q4) so you can report on the original definition of visibility alongside the expanded one. This only works while those prompts stay active: an archived baseline prompt drops out of that report and changes the baseline read, which is one more reason to prune at quarter boundaries only.
• Keep 5 to 10 slots (15 to 30 credits) of headroom so an addition never forces a removal.

---

For agencies: prompts are the scope line

The agencies that make AI visibility profitable sell packages with a fixed shape, not hours and not "whatever the client needs". The cleanest unit to build those packages on is the prompt count: it is what GEOflux meters, it is what the client can understand, and it grows in visible steps. Because each client brand in GEOflux has its own plan and credit wallet, the allocation per client is explicit, which makes it the most defensible line in a proposal.

Map the plan to the tier

Package level and brand size are two different axes. A large brand on Track gets the head topics of its universe; the same brand on the top package gets full coverage. State the prompt allocation in every proposal, and state what is excluded: a package is defined as much by what it leaves out as by what it includes.

Five rules that hold the ladder together

1. Every package includes the report. The monthly numbers are what the client renews for; work without a report is work the client forgets.
2. No fixes before a baseline. Improvement against an unmeasured start cannot be shown, and what cannot be shown gets questioned.
3. Let the data sell the next package. A Track report that shows the client missing from 12 Buy prompts where three competitors appear makes the upgrade obvious without a pitch.
4. Keep strategy out of the entry package, or nobody buys the ones above it.
5. Decide each client's set size at proposal time, and check the wallet before any bulk import.

The repeatable new-client setup

Same list, every client, no exceptions. That is what makes it a product a junior can run.

1. Before kickoff: domain confirmed, competitor set agreed in writing (direct rivals plus adjacent challengers, within the plan's cap), Search Console and analytics access requested when the contract is signed, not at the first meeting.
2. Create the project with the country set to the engagement's scope, not the client's ambitions. The country is chosen once and cannot be changed later, so confirm it first.
3. Load the agreed competitors, within the plan's competitor cap; compare against the brands GEOflux surfaces after the first week and add what is missing as a dated decision.
4. Build the set from the topic tree and the Google data, size it to the plan, import via CSV with Buyer Journey and Brand type filled on every row.
5. Send the approval document; import only when it comes back agreed.
6. Run 30 days, save the baseline as a shared report with a custom date range plus the imported CSV, and record the Non-branded headline from the Overview. Invite the client as a user: on paid and trial projects every seat automatically receives the weekly and monthly summary emails (the only opt-out is the unsubscribe link). If they care about citations, add a watchlist: it tracks source domains you choose (their own domain, key earned-media sites) and sends a daily, weekly or monthly digest of what AI answers cite from them, to the emails you list.

See Agency and User management for the account side.

The audit is your way in

A short visibility check is the strongest opener you have, and GEOflux has a tool built for it: pitch reports. An agency on a paid plan (not a trial) can create a pitch report for a prospect. Each pitch runs on the Growth package: 450 credits (150 prompts), no personas, collecting for 7 days. GEOflux can generate the set for you (25 prompts, 5 topics x 5), or you write your own list, weighted to Evaluate and Buy, in the prospect's language, capped by the 450 credits (150 prompts). You can have up to 10 live pitches at once. A pitch cannot be created for a domain where your agency already has any project, including an existing client. It costs nothing from any client's budget, and if the prospect converts the pitch becomes a paid client project, so the pitch set seeds the paid set. After the week, show them which real questions already name their competitors and not them: concrete, not a deck of claims.

Portfolio hygiene

• One vocabulary across all clients: Buyer Journey (Discover / Evaluate / Buy) and Brand type (Branded / Non-branded) filled on every prompt, plus general tags for product line and baseline quarter, so reports are built the same way every month.
• One headline scope across all clients (last 30 days, all engines, Brand type Non-branded), so a 32% and an 18% are comparable.
• A quarterly portfolio pass: which clients have prompts flat for 90 days, which are near their credit ceiling, which sets have drifted from the ratios.

---

Templates, worked examples, checklist, mistakes to avoid

Prompt templates by type

Fill the brackets in your buyers' language.

Worked example: a 45-prompt Starter set for a UK digital bank

Head volumes are illustrative; 5 slots (15 credits) free.

Check: 11 + 9 + 7 + 8 + 5 + 5 = 45 prompts; Discover 9 (20%), Evaluate 18 (40%), Buy 18 (40%); branded 5 of 45 (11%); 135 of 150 credits, 15 free.

Worked example: a 450-credit Growth set for an online travel agency (shape B, Balanced)

115 prompts across 9 topics (Cheap flights 19, City breaks 17, All-inclusive holidays 16, UK staycations 14, Travel insurance 12, Guided tours 10, Last-minute deals 10, Hotels 9, Comparisons and trust 8), plus 3 personas (family with young kids, couple on a tight budget, business traveller) linked to the same 10 Evaluate prompts where the audience changes the recommendation: 115 + 30 = 145 slots, or 435 credits, with 15 free. The persona slice reads as three separate series against the same 10 questions.

One-page checklist

• Domain, market, language and competitor set (closest rivals plus challengers, within your plan's competitor limit) agreed; project country confirmed before creation.
• Topic tree taken from the site and weighted by commercial value and Google head volume; comparisons and alternatives in their own topic.
• Every prompt has a demand receipt and reads like a real user typed it (8 to 15 words, one intent, buyer's words, market language).
• Ratios at set level: about 20/40/40 Discover / Evaluate / Buy, about 90/10 non-branded; at least 5 prompts per topic.
• Set sized to the plan (15, 50 or 150 prompts; 45, 150 or 450 credits) with 5 to 10 slots of headroom; personas only on the audience-sensitive slice.
• 3 to 5 prompts tested manually per engine; at least one brand appears in the answers.
• CSV built with Buyer Journey and Brand type on every row; import preview shows zero missing values; approval document agreed before import.
• 30-day baseline saved as a shared report with a custom date range plus the imported CSV, with the Non-branded headline recorded from the Overview; headline scope fixed (last 30 days, all engines, Brand type Non-branded).
• Monthly: flag candidates and add dated prompts. Quarterly: archive, re-check ratios and volumes.

Mistakes to avoid