What AI Actually Recommends: a 2,170-Run Study (2026)

Quick answer: Across 2,170 AI-assistant runs (June 8 – July 16, 2026), the sources AI cites most for Shopify inventory, returns and liquidation questions are Reddit (483 citations), the Shopify App Store (453) and YouTube (440) — community and marketplace surfaces, not vendor blogs. Measured against our own domains as a fixed yardstick, grounded engines cite them far more often than Google's AI Overviews do, and the answers are not always current: BULQ, a liquidation marketplace that closed in July 2025, still drew 135 recommendation mentions.

Last updated: July 2026

How we measured this

Forthcast runs a standing measurement corpus: prompts we designed around real Shopify operator questions — inventory forecasting, returns management, surplus liquidation, 3PL selection and supplier sourcing — executed against five AI engines (ChatGPT, Gemini, Perplexity, Google AI Overviews, Claude) with the citations and brand mentions in each answer recorded (runs executed via DataForSEO's AI Optimization endpoints). This study covers the 2,170 error-free runs from June 8 to July 16, 2026, the window with fully validated citation parsing. Two caveats we hold ourselves to: the prompt set is ours, not a random sample of all user queries, and a citation count measures visibility in answers, not product quality. Full methodology lives on our data methodology page.

Disclosure: Forthcast publishes this study and competes in the inventory-forecasting category it measures. Counts below are visibility in our prompt corpus, not endorsements.

Finding 1: AI recommendations run through three platforms

Community and marketplace surfaces are the citation currency — not vendor sites. The most-cited domains in our corpus:

Official Shopify surfaces (shopify.com, its community forum and help center) add up to 695 citations combined. For a merchant, this means the answer an AI gives you is largely assembled from Reddit threads, App Store listings and YouTube videos — with all the freshness and bias trade-offs those carry.

Finding 2: the engines disagree — a lot

Measured against one fixed yardstick — our own domains — grounded engines cite them far more often than Google's AI Overviews do. This is not a measure of how often each engine cites any specific page; we did not measure that. It is the share of runs (same prompts, same window) that cited at least one page on a Forthcast-network domain:

(Claude appeared in only 22 valid runs in this window — too few to compare fairly, so we exclude it from this table.) The gap matters if you publish content: retrieval-grounded engines reward pages they can quote within days, while AI Overviews behaves closer to classic ranking systems and moves on a much slower clock.

Finding 3: who actually gets recommended

A handful of vendors dominate the recommendation surface. The most-mentioned vendor brands in answers to our prompt set, after merging name variants and dropping ambiguous generic tokens:

These are visibility counts in our prompt corpus, not endorsements or market-share figures. Counts are distinct runs in which the brand was mentioned, so a run naming both “B-Stock” and “BStock” counts once. Name variants merged: B-Stock/BStock, Loop Returns/LoopReturns, AfterShip/AfterShip Returns, Inventory Planner by Sage/Inventory-Planner, Return Prime/ReturnPrime, Fabrikator/Fabrikatör. Ambiguous generic tokens (“liquidation”, “streamline”) were dropped.

Finding 4: AI still recommends a marketplace that closed a year ago

BULQ drew 135 recommendation mentions in summer 2026 — it shut down in July 2025. Stocky, which Shopify is sunsetting on August 31, 2026, drew 202. AI answers lag reality by months to a year on vendor status, because the sources they lean on (old threads, old videos, old listicles) lag. If an AI recommends a tool, verify the vendor is alive and the pricing is current before you commit — and if you publish content, keeping vendor-status pages current is exactly the kind of freshness these engines end up quoting.

What we do with this data

This corpus is why our own pages carry dated, source-linked vendor-status and pricing claims — it is measurably what grounded engines pick up. The same dataset powers the demand-forecasting statistics page and the accuracy-correction work across our network. We refresh the runs on a biweekly cycle and will update this study as the next cycles land.