AI Recommendation Index: what six models recommend

The AI Recommendation Index tracks which software brands AI assistants place on a buyer’s shortlist and which one they recommend first.

The current edition combines four category studies run in June 2026. Each study used six buying prompts and the same six-model panel, producing 36 answers per category and 144 answers in total.

This is not a global brand ranking. It is a transparent collection of category snapshots. A CRM answer and a Finnish ecommerce answer do not share the same buyer, language, or competitive set, so collapsing them into one proprietary score would create confidence the data does not support.

The Current Index

The table exposes the central pattern: mention rate alone is not enough. Accounting and CRM both produced ties at the top of share of voice, while the first-pick distribution separated a clear leader.

Download the Q2 2026 dataset

The response-level archive is published under CC BY 4.0. It contains the 144 message outputs, exact prompts, model identifiers, run timestamps, normalized mentions, first-pick proxies, and citations. Provider reasoning, API task IDs, token use, and costs are deliberately excluded.

Required attribution: AI Recommendation Index by Niko Alho, linked to this page. The files are versioned by wave instead of being silently overwritten.

The Method

Each category follows the same high-level design:

1. Define six buying questions covering generic, budget, use-case, and comparison intent.
2. Send the same six prompts to six models with web search enabled.
3. Record every named product, the opening recommendation, and the cited sources.
4. Normalize obvious product-name variants into one brand.
5. Calculate share of voice as answers naming the brand divided by 36.
6. Report the model panel, prompt framing, run month, results, and caveats on the category page.

The panel reported in the category studies was GPT-5.5, GPT-5-nano, Claude Opus 4-8, Claude Haiku 4-5, Gemini 3.1 Pro, and Perplexity Sonar Pro. The point of listing versions is not permanence. It is the opposite: model versions are part of the observation and must be recorded because the next wave may differ.

The downloadable response archive publishes all 24 prompts verbatim and keeps each answer attached to its category, model, market, language, and run timestamp.

Share of Voice Is Not Recommendation Position

AI share of voice answers: was the brand present?

First-pick rate answers: did the answer lead with it?

Those are different commercial positions. A product named in 30 of 36 answers but placed fourth each time has strong shortlist coverage and weak recommendation priority. A product named in 18 answers and placed first in 12 may own a narrower use case more convincingly.

This distinction is why the index does not publish a single blended score yet. Any weighting between coverage and position would be a business judgment, not a fact. The raw components are more useful.

What the Four Studies Already Show

1. High visibility can hide weak positioning

Xero reached 86% share of voice in accounting but opened zero of the 36 answers. Zoho CRM matched HubSpot at 97% share of voice while HubSpot dominated first-pick position. “Mentioned everywhere” and “recommended first” are not synonyms.

2. Prompt design changes the category

Wave’s accounting visibility benefits from the free-software prompt. Linear appears more strongly in project-management prompts aimed at software teams. Notion enters the PM category even though it is not a traditional project-management product. The prompt set defines the market the model is being asked to construct.

3. Language and geography matter

The Finnish ecommerce study surfaced domestic platforms and a Finnish source graph that would not be captured by an English global prompt set. It should not be compared directly with the US-framed accounting study.

4. Cited intermediaries matter

The category studies repeatedly find comparison publishers and specialist editorial sites in the citations. Vendor pages are part of the source graph, but they do not control it. This is where entity authority and digital PR meet LLM SEO.

Limitations

The index is designed to be inspectable, not infallible.

• Small prompt panels. Six prompts can expose a positioning pattern, but they do not represent every buyer question.
• Non-deterministic answers. Live web search and generated output can change between identical runs.
• Model drift. A dated model panel becomes historical as providers update routing and versions.
• Unequal markets. The ecommerce study is Finnish; the others use different framing. Cross-category ranking would be misleading.
• Mention normalization. Product aliases and parent brands require judgment. The public generator contains the normalization patterns, while the export keeps the unedited message text for re-analysis.
• First mention is only a proxy. The first normalized brand mention can be context rather than a positive recommendation. Treat it as a reproducible positional signal, not sentiment analysis.

The dataset makes the classification inspectable, but it does not make a non-deterministic model deterministic. A rerun can verify the method and still produce a different answer.

The Quarterly Protocol

A useful index must show change, not just one snapshot.

For each new wave:

1. Freeze the core prompt set before the run.
2. Keep a small exploratory prompt set separate from the trend panel.
3. Record the exact model identifiers and run dates.
4. Run all brands through the same normalization rules.
5. Publish deltas in share of voice, first-pick rate, and cited-source concentration.
6. Keep old waves available so model drift is visible rather than overwritten.

Do not edit a June result into an August result. Publish a new wave and preserve the baseline.

The stable all-waves manifest points to Q2 2026 now. Q3 and later releases will receive their own directory, response files, summary, and manifest under the same URL structure.

How to Use the Index

If a brand has low share of voice, audit category coverage and the sources retrieved for the missing prompts.

If it has high share of voice but low first-pick rate, the problem is positioning: the web recognizes the brand but consistently frames it as an alternative.

If performance varies heavily by model, inspect which sources each system retrieves and which category language appears around the brand.

Then rerun the same panel after a meaningful release. The measurement loop is detailed in how to measure AI share of voice and tracking AI citations.

The job is not to win a proprietary score. It is to understand where AI-mediated buying research includes the brand, how it positions it, and which source graph creates that answer.