Which accounting software does AI recommend?

QuickBooks and Xero tied on mention coverage in this June 2026 panel: each appeared in 31 of 36 answers. QuickBooks was the first normalized brand in 24 answers, compared with 0 for Xero. The difference describes answer order, not a reviewed preference score.

This category study belongs to the AI Recommendation Index. The September 2026 revision clarifies the original metrics and recomputes the displayed counts from the unchanged Q2 archive.

The mention and first-position table

Mention rate counts a brand once per answer. First position records the first normalized brand occurrence anywhere in the answer text.

The historical dataset calls these fields share of voice and first_pick. Here, first_pick is explicitly a first-mention proxy. A brand mentioned in an introduction may not be the preferred option in the conclusion. Read the comparison of AI share of voice and recommendations for the distinction and denominator rules.

How the model observations differ

The table below separates brand presence from first position for each model’s six answers.

Each cell contains six different questions, not six repeats of the same question. These counts describe variation across the panel. They do not estimate how reliably a model would repeat a recommendation on an identical prompt. Provider names are not proxies for buyer demographics.

The exact buying questions

The prompt set defines what the aggregate represents. The six questions below were each sent to six models.

Named-brand comparisons and brand-free questions are mixed in this historical aggregate. A comparison prompt tests how a model evaluates supplied candidates. A discovery prompt tests which products appear without being supplied. Future studies report those strata separately.

What changes without the named comparison

Excluding q2 removes six answers to a question naming QuickBooks, Xero, FreshBooks, and Wave. QuickBooks, Xero, and Wave each remain at 25/30 mentions. First-mention counts are QuickBooks 19/30, Xero 0/30, and Wave 5/30.

This is a sensitivity check on the existing archive, not a new model run or a semantic reclassification. It supports the narrower finding that equal mention coverage can coexist with different text positions.

The cited source domains

These are counts of citation entries in the normalized Q2 response archive. They are not counts of independent buyers or proof of which source caused a recommendation.

A domain can occur more than once in an answer. A search redirect or grounding URL is retained as recorded; it is not silently attributed to a publisher. Inspect the original URL and its evidence before choosing a content or outreach action.

The citation tracking guide distinguishes cited answers, citation entries, source attribution, and referral visits. A vendor domain being cited does not establish endorsement, and an uncited vendor may still be recommended.

What this study can and cannot tell you

The panel demonstrates differences in brand presence and first-mention position under its declared prompts and model versions. It does not measure product quality, purchase behavior, market share, or revenue.

The responses were collected through the DataForSEO LLM Responses API. Search was requested, but request settings do not prove a search was used for every response. Consumer ChatGPT, Claude, Gemini, and Perplexity experiences may differ.

Brand normalization uses explicit name patterns and can miss an unlisted product or confuse a name with ordinary language. The complete response text is preserved so a reviewer can inspect those cases. The next recommendation layer requires evidence-backed semantic labels with null and tie options.

Do not infer that a brand with zero first mentions was never recommended. Do not infer that a source table identifies a guaranteed placement strategy. The next useful step is to review relevant buyer situations and the answer passages themselves.

Data, attribution, and corrections

Download the Q2 response archive, the spreadsheet-friendly CSV, or the immutable Zenodo release. The original files remain unchanged.

Cite this as AI Recommendation Index by Niko Alho, Q2 2026, with the category, sample of 36 answers, and the first-mention proxy qualification when using positional results. See how to measure AI share of voice for the updated protocol.

Correction, 20 September 2026: replaced recommendation interpretations of first-mention data with explicit proxy labels and regenerated the category and per-model tables from the archived responses. Earlier narrative claims about why a model chose a product were not established by this study and have been removed.