How the AIV Index is measured.

The AIV Index is a measurement instrument, not an optimisation tool. Every design decision in how we run sweeps, score results, and report numbers follows from a single requirement: the index must be an accurate reading of where a brand actually stands in AI-engine responses today.

What we measure

AIV measures share-of-mention — how often a brand appears in AI-engine responses to buyer-intent category prompts, expressed as a proportion of the total mentions captured for that brand and its tracked competitors. We run sweeps across four engines:

• ChatGPT (OpenAI)
• Claude (Anthropic)
• Perplexity
• Gemini (Google)

These are the engines that currently have material adoption in buyer-research workflows. The engine set is reviewed quarterly; engines are added when adoption evidence warrants it.

Prompt design

Each tenant is assigned a set of buyer-intent category prompts — questions that a prospective customer would plausibly ask an AI engine during a research or shortlisting step. Examples: “What are the leading [category] providers in South Africa?” or “Which [category] platforms do accountants typically recommend?” Prompts are drafted during onboarding and reviewed with the tenant before the first sweep runs. They are not changed between sweeps unless the competitive set or product category changes materially.

Prompt stability is deliberate: a consistent prompt set is what makes month-over-month drift meaningful. If we changed prompts between runs, score movements would reflect prompt variation, not genuine changes in AI-engine responses.

How a sweep works

For each prompt, we query each engine N = 3 times (5 for priority prompts and Scale tenants) in independent, stateless requests, with live web search enabled. Each response is then parsed for:

• Brand mention— word-boundary, case-insensitive match for the tenant brand name and any tracked variants.
• List position— ordinal rank if the engine returned an ordered list; unordered presence flag otherwise.
• Citation context— the sentence or clause in which the brand appears, stored verbatim for qualitative review.
• Competitor mentions— identical parsing applied to each tracked competitor in the same response.
• Cited sources— because web search is on, each answer is grounded in live sources; we capture the domains the engine cited so they can be classified (see Reading the sources below).

AI-engine responses are stochastic — the same prompt asked again can return meaningfully different text. We sample each prompt multiple times per engine per cycle (3 by default, 5 for higher-priority prompts and Scale-tier tenants) to smooth this variance enough that a genuine score movement is distinguishable from single-run noise. It does not eliminate variance entirely; it produces a stable-enough rate for directional tracking.

Scoring formula

The index is computed in three steps.

Step 1 — per-engine mention rate. For each brand (tenant and each competitor) on each engine:

rate = responses_with_mention ÷ total_iterations

The rate sits between 0 and 1 at a granularity set by that prompt’s sample count: at the base N = 3 it can be 0, ⅓, ⅔, or 1; at N = 5 (higher-priority prompts and Scale-tier tenants) it can also land on 0.20, 0.40, 0.60, or 0.80. The per-prompt rate is then averaged across all prompts in the tenant’s set — each prompt weighted equally, so a more-sampled prompt buys precision, not extra weight — to produce a single per-engine mention rate for each brand.

Step 2 — per-engine share-of-mention. The tenant’s share on each engine is normalised against the competitive set:

engine_share = brand_rate ÷ (brand_rate + ∑ competitor_rates)

If all tracked brands, including the tenant, have a rate of zero on a given engine, the engine share is recorded as zero (not undefined).

Step 3 — AIV Index. The final index value is the equal-weighted mean of the four per-engine shares:

AIV_Index = (share_ChatGPT + share_Claude + share_Perplexity + share_Gemini) ÷ 4

Equal weighting reflects the current state of the market: no single engine dominates buyer-research workflows in a way that would justify a different weighting. Engine weights are reviewed quarterly. Any change to the weighting scheme will be communicated to tenants before it takes effect and will be noted in the accompanying data release.

Beyond the mention count: conviction & competitor discovery

Share-of-mention tells you if a brand is named. Two further readings — on the Pro and Scale plans — tell you how it is named and who else is in the room. Both come from a single extra pass in which a language model reads each answer we have already stored (the same extractor that powers the public Index), doing two jobs at once:

• Conviction — named versus recommended. For each brand found in an answer, the pass records whether it was merely named in passing or actively recommended. Aggregated across the run, that gives conviction: of the answers that name a brand, how many recommend it, scored per engine. Being named and being chosen are different things, and buyers act on the second.
• Open-capture competitor discovery. The same pass does not only look for brands on the tracked list — it captures every commercial brand the answer names, including untracked ones. Brands that surface often enough, and are not already tracked, become discovery candidates: the rivals the engines are putting next to you that you are not yet watching.

This pass is best-effort and never blocks a reading: if the model call fails, the run still completes with its share-of-mention intact. It runs on the Pro and Scale plans.

Reading the sources

Every grounded answer cites sources, and in most categories the sources — not the brands — decide who gets recommended. For each answer we take the domains the engine cited and sort each into a class: own (the tenant’s site), competitor(a tracked rival’s site), mediator (directories, review sites, comparison pages and rankers), platform (marketplaces and tools that answer in their own right), and other (news, forums, social).

The reading that matters most is the gap: who’s cited instead of you— the mediator and platform sources that appear in answers naming your competitors, where no answer citing them names you. Those are the pages writing your rivals’ story while you are absent from them. Source classification is available on every plan.

The crawler check

An assistant can only recommend a site its crawler is allowed to read, and a surprising number of brands block the AI crawlers at the server without knowing it — one layer below the robots.txt file the usual SEO audit checks. The crawler check fetches a site the way each AI bot does and returns one plain-language verdict per site (Visible, Blocks, Throttled, No server-rendered content, Uneven, or Not tested). It is deliberately conservative: it only calls something a block when the same bot is refused on two separate fetches, so a single flaky refusal never triggers a false alarm. It is free and public — no account needed — and it is the same test the published Index runs.

Sweep cadence

Every tier runs a full sweep monthly, across all four engines. We measure monthly by design: grounded, live-web measurement on four engines is a paid operation on every call, and at our prompt volumes sub-monthly snapshots mostly capture a model’s run-to-run non-determinism rather than real movement in your visibility. A monthly cadence keeps the signal honest and the cost of measurement proportionate.

Need a reading sooner — after a content push, a PR cycle, or a site change — any paid tier can buy an on-demand refresh: a single immediate sweep at your tier’s prompt count, drawn from prepaid credits (R100 / R250 / R500 per run on Starter / Pro / Scale). A refresh reschedules your next included monthly sweep to 30 days out.

Each sweep covers the full prompt set for the tenant. Partial or interrupted sweeps are discarded and re-run; only complete sweep results are published to the dashboard.

What we do not do

AIV is a measurement product. We do not instruct, prompt, or configure any AI engine to prefer or mention a tenant brand. Any technique that would cause an engine to respond differently to measurement sweeps than it would to genuine user queries would corrupt the index and make it worthless as an instrument. It is therefore not only out of scope — it is prohibited, and it would defeat the product’s purpose.

We also do not:

• Manipulate, filter, or discard individual responses that produce an unfavourable result.
• Adjust the competitive set mid-period without tenant consent and a documented reason.
• Report projected or interpolated index values as measured values. If a sweep fails, the dashboard shows a gap, not an estimate.

On zero scores

A score of zero is a true reading. It means the AI engines queried did not mention the tenant brand in any response across the prompt set for that sweep. It is not a product failure. It is the “before” number — the baseline from which subsequent movement is measured. Many tenants begin at or near zero. That is precisely why the measurement is useful.

A note on AI crawling

AI training and retrieval crawlers do not execute JavaScript. A Vercel/MERJ study published in December 2024 confirmed this for current-generation AI crawlers. This has a direct implication for how brands should structure their public content if they want to improve their standing in AI-engine responses: content that exists only in client-rendered markup is effectively invisible to the models being trained on or retrieving from the web. AIV Index scores reflect this reality — we query the engines as a user would, which means our scores capture the result of whatever the engine has indexed, not what a logged-in user sees.

Data handling

Sweep results — prompt text, engine responses, parsed mention data — are stored in a Supabase Postgres database hosted in eu-west-1 (Ireland). This region sits inside the EU adequacy framework that POPIA §72 accepts for cross-border transfers. No customer-of-tenant personal information is stored; the data we hold is limited to the tenant’s brand name, competitor names, prompt sets, and raw engine response text. See the POPIA Notice for the full information-officer disclosure and processing basis.

Questions

Methodology questions can be directed to hello@autoalphaadvisory.co.za. We are happy to walk through the scoring model or discuss how the prompt set for a specific tenant was constructed.