Vanna leads Parse’s AI visibility sample—but this is not a product-quality ranking

Parse has published a useful—but easy to misread—snapshot of how public AI assistants answer text-to-SQL buying questions. On its “Natural language database querying” topic, the company says Vanna appeared in 65% of 149 collected responses, ThoughtSpot in 58%, Snowflake Cortex in 48%, Wren in 40% and Databricks in 39%. The page says the tracked period began July 6 and was updated October 28, 2026.

Those percentages are mention rates, not market share and not answer accuracy. Parse defines “Mentioned in” as the number of collected responses naming a brand divided by all responses collected for that topic. Because one answer can mention several products, the column is not supposed to sum to 100%.

What the split actually shows

The engine-level table is more revealing than the overall ordering. Parse reports Vanna in 46% of sampled ChatGPT answers but 85% of Google AI Mode answers. ThoughtSpot moves the other way: 64% in ChatGPT and 52% in Google AI Mode. Wren is closer across engines, at 38% and 41% respectively.

That divergence is a warning against treating a single aggregate percentage as a stable market verdict. The result depends on which assistant answered, the questions Parse chose, the collection window and whether a brand name was extracted correctly.

Parse’s methodology is unusually explicit about those limits. It says it asks English-language buyer questions with location set to the United States, queries ChatGPT and Google AI Mode, and does not inspect private conversations. It also says its questions are not a measure of buyer demand, repeated answers can share questions and collection dates, and a large response count alone does not establish statistical precision.

How teams should use it

The dataset can help marketing and product teams see which vendors assistants surface when users ask about natural-language database access. It cannot tell a data team which product will generate correct SQL against its schema, enforce its permissions or preserve its business definitions.

For procurement, the right next step is not to select the most-mentioned name. Use the visibility list to build a candidate set, then test finalists on a fixed bank of real questions, verified expected answers, permission boundaries and failure cases. Parse itself makes the same distinction on the topic page: its shortlist is not a claim that one tool has the highest accuracy, and actual performance depends on schema, definitions and configuration.

The headline number is therefore a discovery signal. The more important finding is that the signal changes sharply by assistant—and that Parse gives readers enough methodology to avoid confusing visibility with quality.