AI share of voice vs. recommendations: what the numbers measure

AI share of voice and AI recommendations measure different things. A brand can appear in nearly every answer without being the first brand named. It can also appear first without being the answer’s preferred choice. Measuring those distinctions requires more than a single visibility percentage.

In my Q2 2026 AI Recommendation Index, QuickBooks, Xero, and Wave each appeared in 31 of 36 accounting answers. QuickBooks was the first normalized brand in 24 answers. Wave was first in 5. Xero was first in none.

That is a useful example of equal coverage and different answer positions. It also exposes a limitation in the original measurement: the historical first_pick field is a first-mention proxy. This article keeps the published counts intact and explains precisely what they support.

The accounting example: same visibility, different order

The three accounting brands tie on mention rate and separate on first-mention position.

These are June 2026 observations from six buying prompts across six API models. The panel includes generic, budget, freelancer, ecommerce, startup, and named-comparison questions. It is not a random sample of every accounting question people ask.

The model panel included GPT-5.5, GPT-5-nano, Claude Opus 4.8, Claude Haiku 4.5, Gemini 3.1 Pro, and Perplexity Sonar Pro. The archive records the model identifiers attached to those runs. Listing those models describes the historical panel; it does not imply that current consumer interfaces route users to the same versions.

The response text, prompts, normalized mentions, and citations are available in the Q2 dataset. The dataset is also archived under DOI 10.5281/zenodo.21420271. The counts can be recomputed from saved responses even when a fresh model run produces different answers.

The distinction matters to a reader comparing dashboards. A report showing three identical 86% values has not made an arithmetic mistake. It has summarized one dimension. A second report showing different positions might also be correct. The question is whether each report names the dimension it actually measured.

What AI share of voice measures depends on the formula

The Index’s historical share of voice is an answer-level mention rate. It is not a division of total attention among competitors.

The formula is:

mention rate = successful answers naming the brand / all successful answers in the panel

Count each brand at most once per answer. An answer mentioning QuickBooks ten times contributes one QuickBooks presence, just like an answer mentioning it once. This avoids making a long answer look like ten separate opportunities to appear.

A different measure is competitive mention share:

competitive mention share = the brand's brand-answer mentions / all tracked brand-answer mentions

The denominator now depends on the declared competitor set. If you add Sage, Zoho Books, or FreshBooks to that set, competitive share may change even though QuickBooks appears in exactly the same answers. Mention rate would remain unchanged.

Neither formula should borrow the interpretation of the other. Mention rate describes coverage within the chosen panel. Competitive share describes a relative distribution within the chosen brand set. Neither becomes market share simply because the result is written as a percentage.

For that reason, I now use “mention rate” when precision matters and explain that older Index pages called it SoV. A new label should make the data easier to interpret, not hide a changed calculation. The archived Q2 files retain their original field names.

First mention is not the same as first pick

First mention records where a normalized brand name first occurs. First pick requires the answer to explicitly prefer a product for the stated buyer.

Consider this illustrative answer, which is not a quotation from the study:

QuickBooks is a common option. For a freelancer who bills by the hour, I would choose FreshBooks.

A first-occurrence parser records QuickBooks. A recommendation reviewer records FreshBooks as the preferred option. Both labels can be accurate if they keep their meanings.

Now consider an answer with an alphabetical comparison table. The first row might say FreshBooks, followed by QuickBooks, Wave, and Xero. Row order alone does not establish preference. The conclusion below the table might select Xero, or it might say the choice depends on an unanswered question.

A third answer might recommend QuickBooks for inventory-heavy businesses and Wave for a very small service business without choosing between them. That answer contains conditional recommendations. Assigning a single winner would discard information the model explicitly preserved.

This is why the historical first-pick label needs a qualification. The Q2 first_pick field is reproducible: identify the first matching brand occurrence using the published normalization rules. Its interpretation is limited. It does not classify whether the brand is praised, rejected, quoted, or selected.

A stricter first-pick measure should allow null, ties, and unresolved cases. It should save the supporting passage. Otherwise the scoring system can create a preference that was absent from the answer.

The denominator changes the apparent first-position result

First-position rates across all answers and first-position rates among mentions are different measures.

For QuickBooks, the Q2 accounting panel contains 36 answers, 31 mentions, and 24 first mentions. Dividing 24 by 36 gives 66.7%. Dividing 24 by 31 gives 77.4%. Both calculations use the same observations.

The last row conditions on presence. It can help describe ordering among answers that include the brand, but it hides the answers where the brand did not appear. Report it as conditional first-mention rate rather than silently substituting it for the overall rate.

The same distinction applies to semantic recommendations. A report can ask how often a brand is preferred across all completed answers or how often it is preferred among answers where it is recommended at all. Both may be useful. The numerator, denominator, and eligible population must travel with the number.

Keep answers with no explicit preference in the overall first-pick denominator. Excluding them would make decisive answers look like the entire panel. Keep transport failures out of the denominator and report their count separately, because a missing answer is not a negative recommendation.

Named comparison prompts need their own result

Questions that supply brand names measure comparison behavior rather than spontaneous discovery.

One of the six accounting prompts asks the models to compare QuickBooks, Xero, FreshBooks, and Wave. That is a reasonable buying question, but the candidates have already been supplied. Combining it with brand-free prompts obscures how often those brands enter the answer unaided.

I recalculated the headline comparison without the six answers to that prompt:

The coverage tie and ordering difference remain. This makes the observation less dependent on the named comparison, but it does not remove every limitation. The remaining panel still contains only five questions per model, with no repeated identical observations in Q2.

The sensitivity check also does not convert first mentions into reviewed recommendations. Changing which rows enter the analysis changes the sample. Changing what the labels mean requires a separate classification pass.

For future research, I keep named comparisons as a distinct stratum. They are useful for studying how an answer evaluates known alternatives. Discovery prompts answer the separate question of who gets into consideration without being named by the user.

Buyer context defines the comparison you are measuring

A software category alone is not a complete description of a buyer’s need.

A US business with inventory, a freelancer billing by the hour, and a team needing multiple currencies may all ask about accounting software. Their answers should not necessarily prefer the same product. Treating every difference as model inconsistency can confuse an appropriate response to context with random variation.

The next Index protocol therefore uses paired questions. The base question asks for software for a small US business. Its paired variant adds one constraint, such as a monthly budget or lack of an in-house accountant. The product category, language, and country remain fixed.

Each constraint deserves its own comparison. Changing the budget and country at the same time would make it difficult to attribute the difference to either change. The resulting study can describe how the answer changes under a controlled prompt modification; it still cannot identify every internal reason the model responded differently.

Competitive sets should emerge from the responses. Extract named candidates before mapping known aliases. If an answer introduces a new product, a closed tracking list should not erase it. Preserve the original name and context so a reviewer can decide whether it belongs to the category or is only mentioned as an integration.

This is particularly relevant to CRM. An answer may mention QuickBooks as a HubSpot integration without recommending QuickBooks as CRM software. A brand-name match alone would record presence while missing the role of that presence.

Recommendation stability needs repeated questions

One answer per prompt/model combination cannot establish how consistently a model makes the same recommendation.

Q2 has six questions across six models for each category. That produces 36 answers, but it does not provide six repeats of the same question on each model. A difference between models and a difference between repeated runs are separate forms of variation.

A stability study repeats the exact question under declared conditions, records each response, and compares the labels within that prompt/model cell. If a preferred brand appears in three of five runs, report 3/5. Do not let the small sample masquerade as a precise estimate of every future run.

Fresh requests matter. If later prompts inherit earlier answers, the experiment measures conversation effects as well as the prompt change. Multi-turn buying conversations are worth studying, but they need their own design and should not be mixed with independent single-turn observations.

Timing matters too. Five repeats in one collection window can describe within-wave repeatability. They do not establish week-to-week stability. To measure change over time, preserve the prompt panel, publish dated waves, and record model or surface changes explicitly.

The buyer-context and consistency analyses will be published after collection and review are complete. A planned study is not a new finding. The completed evidence used in this article remains the Q2 archive and the transparent comparison-prompt exclusion.

Citations are another dimension, not a recommendation label

A citation identifies a visible source relationship. It does not necessarily endorse the cited business or establish a preferred product.

A model can recommend Xero while citing an independent accounting comparison. It can cite a vendor’s documentation to explain a limitation. It can cite the AI Recommendation Index as research without recommending Niko Alho as a consultant.

For research publishers, that last distinction matters. A valuable citation may correctly explain a study and credit its author. For software vendors, a valuable observation may be an explicit suitable-option recommendation. They serve different purposes and need separate measurement.

Preserve exact source URLs in the archive, then group them carefully for reporting. Five annotations linking to one page in a single answer are not five independently observed answers. If you count annotations, label the unit. If you count cited answers, deduplicate within the answer.

The AI citation tracking guide sets out separate fields for source links, accurate use, author attribution, and referral visits. Those fields make it possible to inspect whether visibility is producing the kind of recognition the research was intended to earn.

What a useful measurement report should show

A useful report exposes enough detail to challenge its interpretation.

I would expect to see the exact prompts, their selection rationale, the collection surface, model identifiers, dates, completed and failed attempts, full answer text, and the scoring rules. Each recommendation claim should have an evidence passage. Each published percentage should have counts.

The summary should distinguish the following questions:

The first six questions can be investigated with an answer panel. The final question needs evidence beyond that panel. Neither mention rate nor first-pick rate establishes conversion lift, revenue contribution, or market share.

A dashboard becomes easier to trust when it admits what it has not measured. “No semantic review yet” is more useful than a recommendation score built from text order. “Five repeats on six pairs” is more informative than “hundreds of AI conversations” without the sampling structure.

How to use and cite the accounting finding

Use the result as evidence that equal mention coverage can conceal different first-mention positions.

A precise source statement is:

In Niko Alho’s Q2 2026 AI Recommendation Index, QuickBooks, Wave, and Xero each appeared in 31 of 36 accounting answers. QuickBooks was the first normalized brand in 24 answers, Wave in 5, and Xero in none. First mention was a positional proxy, not a reviewed endorsement.

Link to the Index and method or the immutable dataset version. Preserve the June 2026 date and the 36-answer accounting scope when comparing it with newer research.

Avoid translating the result into “AI never recommends Xero” or “QuickBooks wins 67% of buyers.” The study measured neither claim. A finding can be commercially interesting without claiming more than its design supports.

If you are building your own panel, start with the measurement guide. Keep coverage, position, preference, and stability separate. That leaves a report people can inspect and a set of questions the next wave can actually answer.