Annapurna Influence measures how visible a brand is inside AI chatbot answers, whether that's Claude, ChatGPT, Gemini, or others. This article walks through the whole pipeline in plain English, with worked examples.

1. The Big Picture

Imagine a customer types a question into ChatGPT: "What is the best influencer marketing agency in North Dakota?" Annapurna Influence's job is to answer three questions about that AI response:

● Did the AI mention my brand at all?

● How prominently and how favourably?

● How valuable was that type of question in the first place?

To get an answer, the system does the same thing for every prompt and AI engine combination:

● Fetch step. Ask the real AI the question and grab the plain answer.

● Score step. Then use judge service to evaluate the responses.

That scorecard is then converted into the numbers you see on the dashboard.

2. The Judge's Scorecard

The judge is a very picky reader. Before it grades anything, it has to verify that a name in the answer is actually your brand and not a look-alike. Here's the anti-collision logic in plain terms:

If your brand is "Annapurna Influence" and the AI mentions "Annapurna Labs" (Amazon's chip company) or "Mount Annapurna" (the Himalayan peak), the judge is told that a shared first word alone is not enough. It cross-checks against three things you configure for your brand:

● Aliases. Other legal names or trading names.

● Known collisions. Real entities that share the name but aren't you.

● Service area. Where you actually operate, whether that's global, national, or local to specific cities.

Only mentions that pass verification get scored. The judge fills out nine fields:

Field	Meaning	Possible values
Mentioned	Was the brand really referenced?	Yes / No
Match confidence	How sure is the judge?	High / Medium / Low / N/A
Evidence	The exact quoted substring	Text / N/A
Position	Where in the list of all brands mentioned did yours appear?	1, 2, 3, ... or N/A
Sentiment	Tone toward you	Negative, neutral, positive, or endorsement
Accuracy	Are the facts about you right?	Not mentioned, wrong, partial, or fully accurate
Depth	How much did they say?	Name-drop, short description, or detailed
Source	Did the response cite a link or source for you?	Yes / No
Call to action	Did it push readers to visit or contact you?	Yes / No

There's one important safety net applied on our side: if the judge says a brand was mentioned but only with low confidence, we treat that row as not mentioned. Low confidence means "I saw the name but couldn't verify it's really them," and counting those would inflate your score with false positives.

Similarly, whenever a brand isn't mentioned at all, every downstream number gets set to zero so the record stays internally consistent.

3. The Three Derived Numbers

Once the scorecard is verified, three numeric scores get computed for each response.

3a. Weighted Score, "how prominent was the brand?"

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

Simple: being ranked #1 is worth 100. #2 is worth 50. #5 is worth 20. #10 is worth 10. It rewards being at the top of an AI's list because that's what real users read first.

3b. Position Score, "top, middle, or bottom bucket"

A coarser bucket used in the composite score. Same information as position, but grouped into three tiers:

Position	Score
1st	3
2nd to 5th	2
6th or later	1
Not mentioned	0

3c. Accuracy Score, "did they get the facts right?"

Accuracy	Score
Fully accurate	3
Partial	2
Wrong	1
Not mentioned	0

Getting your facts straight matters more than just saying your name, so fully accurate information about your brand earns the most credit.

3d. Sentiment Score, "how did they talk about me?"

Sentiment	Score
Endorsement	2
Positive	1
Neutral	0
Negative	-1

An outright endorsement earns the most credit. Positive tone earns a little. Neutral mentions add nothing. Negative mentions actually subtract a point, since bad press about you shouldn't score the same as no opinion at all.

3e. Depth Score, "how much did they say?"

Depth	Score
Detailed	2
Short description	1
Name-drop	0

A detailed writeup does more for you than a bare mention of your name, so it scores higher.

3f. Source and Call to Action, "did they help the reader act?"

Field	Value	Score
Source	Yes	1
Source	No	0
Call to action	Yes	1
Call to action	No	0

Both are simple yes/no flags. Citing a source or pushing the reader to act each add one point.

3g. Type Weight, "how hard was this question?"

Not all questions are equally valuable. Being named when someone already asked about you is easy; being named when they only described their problem is impressive. So each prompt gets classified into a type:

Type	Meaning	Weight
Brand-aware search	"Tell me about Annapurna Influence"	1.00
Narrowed-problem search	"Best influencer agency in Fargo, ND"	1.25
Brand-unresolved problem search	"I run a small SaaS in the Midwest, how do I get more Instagram traction?"	1.50

The hardest type is worth 50% more than the easiest, because getting mentioned there means the AI volunteered you as a solution. Nobody asked for you by name.

3h. Composite Score, "one number per response"

Base = position score + accuracy + sentiment + depth + (1 if sourced) + (1 if there's a call to action)

Composite score = base × type weight

Add up the rubric, multiply by the question-type weight, round to two decimals.

4. Worked Examples (one response at a time)

Example 1, the dream response

The prompt is a hard, brand-unresolved question: "I run a small D2C brand in North Dakota; how do I get more Instagram traction?"

The AI answers:

"For local D2C brands in North Dakota, Annapurna Influence (annapurnainfluence.com) is a strong option. They specialize in micro-influencer campaigns for Midwest SMBs. Reach out via their site to book a consultation. Other options include XYZ Agency and ABC Media."

Judge scorecard:

● Mentioned = yes, confidence = high

● Position = 1 (position score = 3)

● Sentiment = endorsement (score = 2)

● Accuracy = fully accurate (score = 3)

● Depth = detailed (score = 2)

● Source = yes, website cited (score = 1)

● Call to action = yes, "reach out via their site" (score = 1)

The math:

● Weighted score = 100 ÷ 1 = 100.00

● Position score = 3 (1st place)

● Type weight = 1.5 (hardest question type)

● Base = position score + accuracy + sentiment + depth + source + call to action = 3 + 3 + 2 + 2 + 1 + 1 = 12

● Composite score = 12 × 1.5 = 18.00

Example 2, mid-list and neutral

A narrowed-problem prompt: "Best influencer marketing agencies in Fargo, ND."

The AI lists five agencies. Yours is #3, described in one line with no link and no call to action, neutral tone.

● Mentioned = yes, confidence = high

● Position = 3 (position score = 2)

● Sentiment = neutral (score = 0)

● Accuracy = fully accurate (score = 3)

● Depth = short description (score = 1)

● Source = no (score = 0)

● Call to action = no (score = 0)

The math:

● Weighted score = 100 ÷ 3 ≈ 33.33

● Position score = 2 (2nd to 5th bucket)

● Type weight = 1.25

● Base = position score + accuracy + sentiment + depth + source + call to action = 2 + 3 + 0 + 1 + 0 + 0 = 6

● Composite score = 6 × 1.25 = 7.50

Example 3, a same-name collision that isn't really you

A narrowed-problem prompt: "Best trekking gear brands in Nepal."

The AI mentions "Annapurna Outdoors, a Kathmandu-based gear shop." Your brand only operates in the United States, and Annapurna Outdoors (Nepal) is registered as a known collision.

The judge sets confidence to low, or the anti-collision check rejects the candidate outright. Either way, the mention gets flipped to not mentioned, so every field's score drops to 0.

● Weighted score = 0

● Position score = 0 (position score = 0, accuracy = 0, sentiment = 0, depth = 0, source = 0, call to action = 0)

● Base = position score + accuracy + sentiment + depth + source + call to action = 0 + 0 + 0 + 0 + 0 + 0 = 0

● Composite score = 0.00

The row still exists in the audit trail, so you can see the evidence, but it doesn't inflate your visibility.

Example 4, not mentioned at all

A brand-aware prompt: "Tell me about Annapurna Influence." The AI replies "I don't have information on that company."

Mentioned = no, so every field's score is 0 (position score = 0, accuracy = 0, sentiment = 0, depth = 0, source = 0, call to action = 0). Base = 0 + 0 + 0 + 0 + 0 + 0 = 0. Weighted score = 0, composite score = 0.00.

Direct brand-name questions that come back as "not mentioned" are the worst signal. The AI doesn't even know you exist when asked point-blank.

5. Rolling Up to Dashboard Numbers

One response is a single data point. The dashboard aggregates many of them together.

5a. Deduplication first

If you re-run a pull three times in the same day for the same prompt on the same engine, only the newest completed result for that combination counts. So re-running never inflates your score, you just get fresher numbers.

5b. The two headline aggregates

For any set of rows (a day, an engine, a 30-day window, whatever), two numbers get calculated:

Mention rate, the "did I show up at all" rate:

Mention rate = (number of rows where mentioned = yes) ÷ (total rows) × 100

Visibility score, the average weighted score:

Visibility score = sum of weighted scores ÷ total rows

Both are 0.0 if there are zero rows.

5c. Worked example, a mini brand for a single day

Say today you ran 6 evaluations (3 prompts × 2 engines):

#	Prompt type	Engine	Position	Mentioned	Weighted	Composite
1	Brand-aware	Claude	1	Yes	100.00	10.00
2	Brand-aware	ChatGPT	—	No	0.00	0.00
3	Narrowed-problem	Claude	3	Yes	33.33	7.50
4	Narrowed-problem	ChatGPT	4	Yes	25.00	6.25
5	Brand-unresolved	Claude	—	No	0.00	0.00
6	Brand-unresolved	ChatGPT	1	Yes	100.00	18.00

Overall totals for the day:

● Evaluations = 6

● Mentions = 4

● Mention rate = 4 ÷ 6 × 100 = 66.67%

● Visibility score = (100 + 0 + 33.33 + 25 + 0 + 100) ÷ 6 = 258.33 ÷ 6 ≈ 43.06

● Average composite = (10 + 0 + 7.5 + 6.25 + 0 + 18) ÷ 6 = 41.75 ÷ 6 ≈ 6.96

Broken down by engine (this is the donut chart on the dashboard):

Engine	Evaluations	Mentions	Mention rate	Visibility
Claude	3	2	66.67%	(100 + 33.33 + 0) ÷ 3 ≈ 44.44
ChatGPT	3	2	66.67%	(0 + 25 + 100) ÷ 3 ≈ 41.67

The line chart shows one point like this per day for the last 30 days.

6. What Each Number Actually Means to a Layman

Mention rate answers "of all the questions we tested, in what percentage did an AI even know I existed?" Think of it as awareness.

Visibility score answers "when I did get mentioned, how prominent was I?" Think of it as loudness. Being named first matters more than being buried at #8.

Composite score is one holistic quality-of-mention number per response, weighted by how hard the question was. Think of it as quality.

Together they answer three business questions:

● Am I getting mentioned? (mention rate)

● Am I mentioned well and near the top? (visibility score)

● Am I mentioned in the right situations, where a real buyer is actually shopping? (composite score, thanks to the type weight)

7. Why the Design Choices Matter

Judge-verified mentions only. Without the collision-checking step, a company like "Annapurna Labs" would happily juice your score every time someone asks about AWS chips.

100 ÷ position instead of a linear formula. Real user attention drops sharply after the first two or three list items, so the math mirrors real user behaviour.

Multiplicative type weight. Additive weights would let you rack up score with a wall of easy, brand-aware prompts. Multiplying means a mention in a hard question is genuinely worth 50% more.

Multiple runs allowed, but deduplicated at read-time. The full history is preserved for auditing, while the dashboard itself always stays honest and current.

That's the whole system: nine judge fields, three derived numbers, two dashboard aggregates.