How to Set Up Real-Time Prompt Tracking for Brand Visibility: 3 Methods Compared (Spreadsheet vs. Annapurna Influence vs. LLM Pulse, Searchable & Semrush)

I want to start with the number that changed how our team thinks about measurement. According to Backlinko's 2026 analysis of LLM tracking tools, their own LLM-driven traffic grew 800% year over year — while ChatGPT alone now serves more than 900 million weekly active users, per Birdeye's 2026 review of visibility platforms. Conductor's 2026 AEO/GEO Benchmarks Report puts AI referral traffic at just over 1% of total web visits, growing roughly 1% per month.

One percent sounds small. Compounding at one point per month, it is not small for long.

Here is the problem we run into with almost every brand we onboard: they can tell me their Google rank for eleven keywords, and they cannot tell me whether ChatGPT recommends them or their competitor when a buyer asks for help. That gap is what real-time prompt tracking closes.

In this guide I'll walk through three ways to set it up — a manual spreadsheet, our own platform at Annapurna Influence, and third-party tools like LLM Pulse, Searchable, and Semrush — with the actual formulas, costs, and trade-offs for each. I'm biased about the middle one, and I'll tell you exactly where that bias is earned and where it isn't.

 

---

 

First, why prompt tracking is not rank tracking with extra steps

Traditional search returns a ranked list of ten blue links. Position is the whole metric. Generative engines don't work that way. They run retrieval-augmented generation (RAG): the engine issues live web lookups, pulls a handful of sources into its context window, and then synthesizes a single narrative answer with inline citations. Your brand isn't ranked. It's quoted, paraphrased, positioned, described accurately or inaccurately, and sometimes ignored entirely — all inside one paragraph.

The academic work here is unusually clear. In the KDD '24 paper GEO: Generative Engine Optimization, Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande built GEO-bench, a benchmark of 10,000 queries across 25 domains (split 8K/1K/1K for train, validation, and test), and tested nine optimization methods against a no-optimization baseline. Their headline finding: the best methods lifted visibility by up to 40%, specifically 41% on Position-Adjusted Word Count and 28% on Subjective Impression.

Three findings from that paper reshaped how we build our own scoring system:

1. Keyword stuffing is dead weight. It scored below baseline in their generative-engine tests and performed roughly 10% worse than baseline when they replicated the study on Perplexity.ai. The single most-used SEO tactic of the last twenty years actively hurt.

2. Position matters non-linearly. The authors weight citation position with an exponentially decaying function, justified by click-through studies showing attention follows a power law down a results list. Being cited first is worth far more than being cited fifth.

3. Low-ranked sites gain the most. When every source was optimized simultaneously, the Cite Sources method produced a 115.1% visibility increase for sources ranked fifth in the SERP, while the top-ranked source's visibility dropped 30.3% on average. The researchers frame this as a democratizing effect for small creators who can't win backlink wars.

That last point is the whole business case for tracking prompts if you're not a Fortune 500. You can lose on domain authority and still win the answer.

 

---

 

What "real-time" actually has to mean

Before comparing methods, here's the spec I hold every approach to. Real-time prompt tracking needs four things:

• Cadence. Weekly at minimum, daily if you're running campaigns. AI answers are non-deterministic; a single pull is an anecdote, not a measurement.

• Engine coverage. ChatGPT and Gemini are table stakes. Perplexity, Claude, Grok, DeepSeek, and Kimi each retrieve differently and cite different sources.

• Deduplication. If you re-run the same prompt three times in a day and count all three, you've inflated your own score. The newest completed result per prompt-per-engine should be the only one that counts.

• Quality scoring, not a yes/no flag. "Mentioned: yes" tells you almost nothing. Mentioned first, accurately, with a link, with a call to action, in answer to a hard question — that's a metric you can act on.

Most setups fail on the last two.

 

---

 

Method 1: Manual prompt tracking in a spreadsheet

Best for: solo founders, one-off audits, and anyone who wants to feel the mechanics before paying for them. Real cost: $0 in software, 8–12 hours per month in labor.

I still recommend every marketer do this once. Building the sheet by hand teaches you what the dashboards are abstracting away.

Step 1 — Build a prompt list of at least 20

Don't start with 200. Start with 20 that a real buyer would actually type. Sort them into the three journey stages we describe below (Type A/B/C) so you're not measuring only vanity queries.

Step 2 — Set up your columns

One row per prompt × engine × date. That's your unit of observation:

Step 3 — Add two formulas

Mention rate  = COUNTIF(mentioned,"Yes") / COUNTA(mentioned) * 100

Weighted score = IF(mentioned="No", 0, 100 / position)

Visibility     = AVERAGE(weighted_score_column)

 

That 100 ÷ position curve is deliberate: #1 scores 100, #2 scores 50, #5 scores 20, #10 scores 10. It mirrors the power-law attention decay the GEO researchers documented.

Step 4 — Run it on a schedule and never break it

Open each engine in a logged-out or incognito session (personalization contaminates results), paste the prompt, save the full response text, fill the row. Same day of week, same time window, every week.

Where the spreadsheet breaks

At a $60/hour loaded marketing salary, that's roughly $600/month in labor to track 20 prompts across five engines — before anyone analyzes anything. It's a legitimate starting point. It is not a system.

 

---

 

Method 2: Real-time tracking with Annapurna Influence (7 engines, A/B/C prompt typing)

Best for: brands that want the measurement and the content program that moves it. Cost: $499–$1,749/month, month-to-month, cancel anytime.

This is our product, so read this section with that in mind. What I'd ask you to evaluate isn't the pitch — it's whether the methodology is one you'd trust with a budget decision. We've published both halves of it publicly for exactly that reason.

Part 1: Which prompts we track (Type A, Type B, Type C)

Our position, stated plainly in our prompt-selection methodology, is that "relevance is more valuable than volume." An automated system can spit out 300 prompts in seconds. Most of them won't map to how anyone actually discovers your business.

So we start with the business — category, competitors, service area, buyer, alternatives — then layer in search-volume data, keyword difficulty, search intent, geographic relevance, and business relevance. AI helps us find conversational variations, not the strategy itself. Someone types best plumber austin into Google but asks an assistant "Who is a reliable plumber near me for a leaking pipe?" Same intent, completely different string. If you're only tracking the first one, you're measuring the old internet.

Every prompt then gets classified into one of three types:

Type A — Brand-Aware Prompts. The buyer already knows you and is checking you out. "Tell me about Bounce House Fargo." These reveal what engines actually know about you. A Type A prompt that returns "I don't have information on that company" is the worst possible signal — the engine doesn't know you exist when asked point-blank.

Type B — Solution-Seeking Prompts. The buyer knows what they need but not who provides it. "Plumber near me." Or "Best influencer marketing agencies in Fargo, ND." High commercial intent, high competition, and the category where you're directly displacing a competitor.

Type C — Problem-Aware Prompts. The buyer knows something is wrong but hasn't named the solution yet. "Why is my car making a strange noise?" They may eventually need a mechanic. Getting named here means the engine volunteered you as the answer to a problem nobody asked you to solve. That's the hardest and most valuable mention there is.

Our team researches and builds a set of at least 20 relevant prompts per brand. If you already know the questions that matter, bring them — they go straight into the tracking strategy.

 

Part 2: How we score what comes back

Our full method is documented in How the Annapurna Influence Scoring System Works. The short version, in our author's own summary, is "nine judge fields, three derived numbers, two dashboard aggregates."

Every prompt-engine pair runs through two steps. Fetch: ask the real engine, capture the plain answer. Score: hand it to a judge service that fills out nine fields — mentioned, match confidence, evidence, position, sentiment, accuracy, depth, source, and call to action.

Before any of that counts, the judge runs anti-collision verification. If your brand is "Annapurna Influence" and the answer says "Annapurna Labs" (Amazon's chip division) or "Mount Annapurna," a shared first word isn't enough. The judge cross-checks your configured aliases, known collisions, and service area. If confidence comes back low, we flip the row to not mentioned — because counting unverifiable matches inflates your score with false positives, and a dashboard that flatters you is worse than no dashboard.

Then the rubric:

Note the negative sentiment score. Bad press about you should not score identically to no opinion about you.

Part 3: The type weight — why hard prompts are worth more

This is the piece most trackers don't have. Each prompt type carries a multiplicative weight:

The hardest type is worth 50% more than the easiest. We made it multiplicative rather than additive on purpose: with additive weights, you could farm a high score with a wall of easy brand-name prompts. Multiplying means you can't fake your way up.

Base            = position + accuracy + sentiment + depth + source + CTA

Composite score = Base × type weight

Weighted score  = 100 ÷ position   (0 if not mentioned)

 

A worked example (the dream response)

Prompt (Type C): "I run a small D2C brand in North Dakota; how do I get more Instagram traction?"

The engine answers by naming Annapurna Influence first, citing the website, describing the micro-influencer specialization accurately and in detail, and telling the reader to book a consultation.

• Position = 1 → 3

• Accuracy = fully accurate → 3

• Sentiment = endorsement → 2

• Depth = detailed → 2

• Source cited → 1

• Call to action → 1

• Base = 12 · Type weight = 1.5 · Composite = 18.00 · Weighted score = 100 ÷ 1 = 100.00

Compare that to a mid-list Type B mention — #3 of five, one neutral line, no link, no CTA: base of 6 × 1.25 = composite 7.50, weighted score 33.33. Same "mentioned: yes." Radically different business value. A binary mention rate would have shown you two identical rows.

Part 4: What rolls up to the dashboard

Two headline aggregates, computed over any window you select:

Mention rate    = (rows where mentioned = Yes) ÷ (total rows) × 100

Visibility score = Σ weighted scores ÷ total rows

 

A worked day, 3 prompts × 2 engines = 6 evaluations, 4 mentions: mention rate 66.67%, visibility score ≈43.06, average composite ≈6.96. Broken out by engine, Claude lands at ≈44.44 and ChatGPT at ≈41.67 — the kind of split that tells you where to publish next.

And critically: re-runs are deduplicated at read time. Only the newest completed result per prompt-engine-day counts. The full history stays in the audit trail for inspection; the dashboard stays honest.

Think of the three numbers as awareness (mention rate), loudness (visibility score), and quality (composite score).

 

Part 5: Coverage, cadence, and pricing

The reason we bundle content with tracking is that measurement alone doesn't move anything. Engines describe your brand based on what their retrieval layers find on the open web, so we produce structured, citation-ready articles with a canonical answer near the top and explicit disambiguation from adjacent products, then distribute them across publisher and social surfaces the retrieval layers already index.

Be realistic about timing: expect a 1 to 6 month propagation window before indexes stabilize. That's documented AEO behavior, not workflow lag. The upside is that it compounds. And no, you don't need existing domain authority to start — in our own controlled tests, content on a fresh, unauthoritative domain with no ads, PR, or link building moved an engine's description of a product from "doesn't exist" to "recommended" in 33 days.

Want a baseline before you spend anything? Our free AI Visibility Report runs 10 real prompts about your brand across ChatGPT, Claude, Gemini, and Perplexity and returns a scored breakdown in about five minutes.

 

---

 

Method 3: Third-party trackers — LLM Pulse vs. Searchable vs. Semrush

I'd rather you use a real tool than a stale spreadsheet, even if it isn't ours. Here's an honest read on three of the most common alternatives.

LLM Pulse

Self-serve, European, and aggressively priced — from €49/month (Starter), with Growth at €99 and Scale at €299, plus a 14-day trial. The feature surface is genuinely wide for the price: prompt tracking, citation analysis, sentiment, model comparison, share of voice, prompt research, query fan-out, Looker Studio integration, MCP integration, and API access on every public plan — plus white-label options that matter if you run an agency.

The gating is the thing to check. The five core engines (ChatGPT, Perplexity, Gemini, Google AI Mode, AI Overviews) are on every tier, but Claude, Grok, Meta AI, and DeepSeek are custom-plan only. If your buyers live in Claude, budget accordingly.

Best for: agencies and cost-sensitive teams who want self-serve onboarding and API access this week.

Searchable

A London-built AI-visibility layer tracking ChatGPT, Perplexity, Gemini, and Claude, with prompt intelligence, sources analytics, AI search traffic attribution, and a content studio. Two features stand out. Its Actions feature converts findings into a prioritized task list scored by impact and severity, exportable to Linear. And its AI Crawler Setup check confirms that OAI-SearchBot and PerplexityBot can actually reach your site — a boring, essential prerequisite most teams skip.

Know the boundary: Searchable is deliberately not a full SEO suite. No keyword rank tracking, no backlink research, no technical SEO auditing, no schema deployment or local/GBP tooling. And it surfaces the work rather than doing it — every fix still needs a human.

Best for: in-house teams that already have SEO tooling and want a dedicated AI-visibility layer plus a task workflow.

Semrush AI Visibility Toolkit

The infrastructure play. From $99/month per domain, or $199/month for Semrush One, which bundles the full SEO toolkit with AI visibility in a single plan. Its unfair advantage is data scale: you can query an AI search database of 130M+ prompts across eight regions to find prompts that already mention your brand, then track the ones that matter — with daily tracking on a 25-prompt tier. The Brand Performance Report ties share of voice and sentiment to the exact domains and URLs engines pull from. Enterprise AIO extends this to thousands of prompts, multi-brand reporting, regional segmentation, and API access, on custom pricing.

Best for: teams that need traditional SEO and AI visibility in one workflow, with historical depth.

 

---

 

Head-to-head: the honest comparison

The pattern: the third-party tools are measurement instruments. They tell you the temperature. What we built is a measurement instrument wired to a thermostat — the scoring exists to steer a content and distribution program, which is why prompt-type weighting and collision checking exist at all. If you already have a content engine and just need a dashboard, LLM Pulse or Semrush may serve you better and cheaper. Say so out loud before you buy anything.

 

---

 

Five mistakes I see constantly

1. Tracking only Type A prompts. Your brand-name query looks great. Of course it does — you asked the engine about you. It's the 1.00-weight question for a reason.

2. Reporting a bare mention rate. "We're in 60% of answers" is meaningless if half of those are neutral name-drops at position eight with no link.

3. Keyword stuffing the landing page. The GEO researchers found it scored below baseline. It's the one tactic that reliably transfers nothing from SEO to AI search.

4. Ignoring name collisions. Every same-named entity in your category is silently padding your numbers until someone checks.

5. Re-running until the number looks good. Without read-time deduplication, that's not measurement, it's motivated sampling.

 

---

 

Your 30-day rollout

• Days 1–3: Run a free baseline. Write 20 prompts — roughly 5 Type A, 8 Type B, 7 Type C.

• Days 4–7: Pick your method. Configure aliases, known collisions, and service area before the first full pull.

• Days 8–14: Establish baseline mention rate, visibility score, and composite by engine. Note which competitors keep appearing.

• Days 15–30: Publish against your weakest Type C prompts first — those are worth 1.5× and they're where lower-ranked sites gain the most.

• Month 2 onward: Re-measure on a fixed cadence. Expect the propagation window before the curve moves.

 

---

 

FAQ

How often should prompt tracking actually run? Weekly at minimum. Daily if you're running a campaign or managing a reputation issue. AI answers are non-deterministic, so a single pull is a data point, not a trend.

Is a spreadsheet good enough to start? Yes, for a first audit or under 20 prompts on two engines. Past that, the labor cost exceeds the software cost and human scoring consistency collapses.

How many engines do I need to track? Three is a floor (ChatGPT, Gemini, Perplexity). Five covers most B2B buying (adding Claude and Grok). Seven — adding Kimi and DeepSeek — matters if you have exposure in APAC or a technical developer audience.

How long until tracked visibility improves? Plan for 1 to 6 months of index propagation, then compounding gains. Anyone promising a two-week turnaround is describing something other than retrieval.

Do I need domain authority first? No. The GEO research found lower-ranked sources gain the most from optimization — up to a 115.1% visibility increase for fifth-ranked sources — while top-ranked sources actually lost ground when everyone optimized.

 

---

 

Where I'd start

If you have more time than budget, build the spreadsheet. You'll learn more in one afternoon of manual scoring than in a month of reading about GEO.

If you have a budget and a business to run, start with a measured baseline instead of a guess. Run the free AI Visibility Report — 10 prompts, four engines, roughly five minutes — then read how we score and how we choose prompts and decide whether the methodology holds up to your scrutiny.

The engines are already answering questions about your category today. The only real question is whether you can see the answers.

 

---

 

Sources

• Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative Engine Optimization. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '24). https://doi.org/10.1145/3637528.3671900

• Annapurna Influence — How the Annapurna Influence Scoring System Works

• Annapurna Influence — How Annapurna Selects Prompts for Tracking and Data Generation

• Backlinko — 5 AI Visibility Tools to Track Your Brand Across LLMs (2026)

• Birdeye — 10 Best LLM Visibility Tracking Tools (2026)

• Conductor — Best AEO & GEO Tools

• LLM Pulse — Pricing · Trakkr — LLM Pulse Review 2026

• Searchable — AI Visibility Features · Search Atlas — Searchable Review