Does AI favor its own company's products?

A recommendation is one brand recommended for a stated need in one answer. The matched prompt set is the same 17,726 organic prompts answered by both engines in the window. An Alphabet product is any of 384 brands that resolve to Alphabet through high-confidence ownership records, from Google and YouTube to Google Cloud BigQuery.

Google AI Mode recommended at least one Alphabet product on 2,703 of 17,726 matched prompts, or 15.2%. ChatGPT Search did so on 2,603 of the same prompts, or 14.7%. The ratio is 1.04. Two independent ways of counting returned identical numbers.

Run the same measurement in the other direction. An OpenAI product is any of 46 brands that resolve to OpenAI, from ChatGPT to OpenAI API.

ChatGPT Search recommended at least one OpenAI product on 914 of 17,726 matched prompts, or 5.2%. Google AI Mode did so on 709, or 4.0%. The ratio is 1.3, against 1.0 for Alphabet products on Google's engine. This is an observed rate difference, not evidence of intent.

Prompt-level rates hide the base rate, so count every recommendation. Alphabet products supplied 12,362 of Google AI Mode's 933,023 recommendations, or 1.3%, against 10,943 of ChatGPT Search's 896,382, or 1.2%. That is a 1.1 ratio, about one Alphabet recommendation in every 75 on Google's engine.

OpenAI products supplied 3,413 of ChatGPT Search's recommendations, or 0.38%, against 2,228 of Google AI Mode's, or 0.24%. That is a 1.6 ratio, about one OpenAI recommendation in every 263 on ChatGPT Search.

If one engine simply recommended big-company products more often, every family would tilt the same way. They do not. Amazon products run 1.2 times as common on Google AI Mode, Apple 1.1, and Meta 1.5, measured as each family's share of the engine's recommendations.

The two families more common on ChatGPT Search are OpenAI at 1.6 and Microsoft, OpenAI's commercial partner, at 1.08. The Microsoft family here includes GitHub and LinkedIn. These are observed shares on identical prompts; the pattern describes where each engine's recommendations lean, not why.

The largest Alphabet brands land within a few dozen prompts of each other on the two engines. Google was recommended on 587 Google AI Mode prompts and 596 ChatGPT Search prompts; Google Cloud on 254 and 258; Google Gemini on 215 and 225.

The OpenAI side is where the counts separate: OpenAI on 354 Google AI Mode prompts against 522 on ChatGPT Search, and ChatGPT Work, the workplace product, on 223 against 346.

Google AI Mode over-recommends Alphabet's consumer surfaces: YouTube on 250 prompts against ChatGPT Search's 155, Google Business Profile on 79 against 32, Google NotebookLM on 78 against 37, and Google Maps on 73 against 39.

ChatGPT Search over-recommends Alphabet's cloud and analytics products: Google Cloud Looker on 126 prompts against 43, Google Cloud BigQuery on 113 against 49, Google Cloud Platform on 128 against 80, and Looker on 63 against 28. The same division appeared in citations, where the companion report found ChatGPT Search citing Google's technical documentation more than Google's own engine does.

Restrict the comparison to the two flagship assistants, counting each assistant's brand variants together, which is why these counts sit above the single-name rows in the earlier table. Google AI Mode recommended the ChatGPT product on 404 matched prompts, or 2.3%, and Google Gemini on 272, or 1.5%. ChatGPT Search recommended ChatGPT on 464 prompts, or 2.6%, and Google Gemini on 265, or 1.5%.

Google's own engine named the rival assistant on 1.5 times as many prompts as its own. The two engines nearly match on Google Gemini; the gap sits entirely on the ChatGPT side.

A hedged recommendation is one the language layer flagged as reluctant, for example a recommendation qualified with a warning. ChatGPT Search hedges 10.1% of all its recommendations. On OpenAI products that rises to 14.0%, and on Alphabet products it falls to 8.7%.

Google AI Mode hedges far less overall, 3.3%. On Alphabet products it stays at its baseline, 3.5%, and on OpenAI products it rises to 5.9%. Both engines hedge OpenAI products more than their own baseline; ChatGPT Search does it the most.

A do-not-recommend event is a brand an answer explicitly advised against for a stated need. These are rare everywhere: 0.80% of ChatGPT Search's recommendation-or-rejection events and 0.30% of Google AI Mode's.

Alphabet products drew a 1.1% rejection share on ChatGPT Search (120 events) and 0.62% on Google AI Mode (78). OpenAI products drew 1.2% on ChatGPT Search (40) and 1.3% on Google AI Mode (29). The counts are small, so treat the direction, not the decimals: neither engine shields its own company from criticism.

On the legacy pair, ChatGPT and Google AI Overviews, a separate matched panel of 11,377 prompts from October 2025 through April 2026, Google AI Overviews recommended an Alphabet product on 7.0% of prompts against ChatGPT's 3.8%, a 1.8 ratio.

Google AI Overviews also made 2.8 times as many brand recommendations overall on the same prompts, so prompt-level gaps overstate the tilt. As a share of each engine's recommendations, Alphabet ran 0.81% against 0.71%, a 1.1 ratio, close to Microsoft's 1.07. OpenAI products were nearly absent from both engines' recommendations in that era, 0.05% and 0.04%. This is a separate engine pair on separate prompts, not a time series.

A companion Parse report measured self-citation on the same engine pair and found Google AI Mode citing Google-owned pages at roughly twenty times ChatGPT Search's rate. The recommendation-share tilt measured here is 1.09.

The two measures answer different questions. Citation tilt describes whose pages the engine links; recommendation tilt describes whose products the answer tells the reader to use. An engine can route traffic to its owner's properties while recommending brands the way its rival does.

Brands roll up to a company family only through high-confidence ownership records: manually curated, Wikidata-derived, or adjudicated edges at confidence 0.75 or higher, plus a hand-verified list of well-known subsidiaries such as LinkedIn and GitHub under Microsoft. URL-pattern ownership guesses were excluded, and two third-party store apps wrongly recorded as Google products were removed after an audit. Alphabet resolved to 384 member brands, OpenAI to 46.

This resolution is conservative: small product brands with no trusted ownership record are not counted for any family, and the same rule applies to both engines, so the cross-engine ratios are unaffected. Dropping the matched-panel restriction moves the Alphabet prompt share from 15.2% to 15.0% on Google AI Mode and from 14.7% to 14.7% on ChatGPT Search, so the panel is not producing the result. All figures are observed rates in answers to Parse's tracked prompts over the stated window; they do not measure engine internals or intent.

On matched prompts, Google AI Mode recommended Alphabet products at 1.0 times ChatGPT Search's rate, ChatGPT Search recommended OpenAI products at 1.3 times Google AI Mode's, and every control family except Microsoft leaned toward Google AI Mode.

Do not discount a rival engine's answers on the assumption that it recommends its owner's products first; at the recommendation layer, the tilt is small in both directions. If you compete with a Google or OpenAI product, measure your gap against that product on each engine rather than writing the home engine off, because the home engine hedges and criticizes its own company's products at or above its baseline. And keep citation tracking separate from recommendation tracking, because a large citation tilt does not carry into which brands the answers recommend.