ChatGPT Recommendation Algorithm Changes: September 2026
Last updated: October 8, 2026
Published September 8, 2026
Our analysts ran a rolling study of how ChatGPT decides which companies and products to recommend when someone asks it for a recommendation. The most recent wave, fielded from July 14 to August 28, 2026, consisted of 47,300 commercial prompts across 62 purchase categories. Every prompt was run in a clean session with memory and personalization disabled, and every answer was scored for which brands it recommended rather than merely mentioned.
A word on what is being measured, because the distinction drives everything that follows. This report models commercial recommendations, meaning the moment ChatGPT puts forward a specific company or product as an option worth considering. That is a separate behavior from citation, which is the act of linking to a source in support of a factual claim. The two overlap far less than the industry assumes: across our commercial prompt set, only 23.8% of cited domains belonged to a brand the same answer went on to recommend. A brand can be the most-cited source in a category and still never be recommended as an option in it.
The weights below come from a regression of brand-level features against recommendation frequency, refit each quarter against a prompt set that we hold constant between waves. Holding the prompts fixed is what allows the quarter-over-quarter movement to be read as a change in the model's behavior rather than a change in our sampling. In the sections below, we publish the current algorithm, the eleven quarters of drift that produced it, the breakdowns by purchase category and prompt type, and the specific product releases that moved individual factors.
The Factors Behind ChatGPT's Commercial Recommendations
In the table below, we break down the nine features that, together, explain the recommendation outcome in our model, along with how far each has moved since our first wave in Q1 2024. Weights are expressed as a share of total explained variance and sum to 100%.
As a check on our own work, we compared the result against a separately published breakdown from the SEO agency First Page Sage, the only other public attempt we found to weight ChatGPT's commercial recommendations rather than its citations. Their model puts authoritative list mentions at 41% of the outcome and online reviews at 16%. Ours, built on a different factor taxonomy and a different prompt set, puts them at 38.6% and 15.9%. Two independent studies converging on the same leading factor carries more weight than either one alone.
| Factor | Weight | Change vs. Q1 2024 (pp) | Direction |
|---|---|---|---|
| Authoritative list mentions | 38.6% | +22.4 | Rising |
| Online reviews and rating aggregates | 15.9% | +7.8 | Rising |
| Awards, accreditations, and affiliations | 14.2% | +7.9 | Rising |
| Customer examples and usage data | 11.4% | +6.0 | Rising |
| Social and community sentiment | 9.7% | +6.5 | Rising |
| Brand familiarity carried in model weights | 4.8% | -19.3 | Declining |
| Search index position on fanout queries | 3.1% | -15.6 | Declining |
| Owned-site content and machine-readability | 1.6% | -8.0 | Declining |
| Backlink authority | 0.7% | -7.7 | Declining |
- We found that one factor, appearances on independent ranked lists, now carries more weight than the four legacy web factors combined, which together account for 10.2% of the model.
- Our data showed that the five reputational factors, meaning list mentions, reviews, awards, customer examples, and sentiment, hold 89.8% of the weight, up from 39.2% in Q1 2024.
- We found that brands absent from every third-party ranked list in their category appeared in only 6.1% of recommendation sets, irrespective of how thorough their own website was.
ChatGPT Recommendation Factor Weights, Q1 2024 to Q3 2026
The speed at which these weights moved is the more useful part of the dataset. In the table below, we show every factor at six-month intervals across eleven quarters. The table is followed by two charts that separate factors losing influence from those gaining it.
| Factor | Q1 2024 | Q3 2024 | Q1 2025 | Q3 2025 | Q1 2026 | Q3 2026 |
|---|---|---|---|---|---|---|
| Authoritative list mentions | 16.2% | 20.8% | 26.3% | 30.1% | 35.4% | 38.6% |
| Online reviews | 8.1% | 9.4% | 11.2% | 13.4% | 14.1% | 15.9% |
| Awards and accreditations | 6.3% | 7.9% | 9.8% | 11.4% | 12.9% | 14.2% |
| Customer examples | 5.4% | 6.1% | 7.6% | 8.9% | 10.2% | 11.4% |
| Social sentiment | 3.2% | 3.9% | 5.1% | 6.8% | 8.3% | 9.7% |
| Brand familiarity in weights | 24.1% | 22.1% | 13.4% | 12.4% | 7.9% | 4.8% |
| Search index position | 18.7% | 15.9% | 17.4% | 9.8% | 5.6% | 3.1% |
| Owned-site content and markup | 9.6% | 8.3% | 5.5% | 4.7% | 3.2% | 1.6% |
| Backlink authority | 8.4% | 5.6% | 3.7% | 2.5% | 2.4% | 0.7% |
Two features of these curves are worth reading carefully. The first is the single non-monotonic line in the dataset: search index position rose between Q3 2024 and Q1 2025, from 15.9% to 17.4%, before resuming its decline. That bump lines up with the launch of ChatGPT search in October 2024, when the assistant began grounding a much larger portion of answers in a live index. The lift did not hold. As retrieval matured and the model began reading further down the result set and across more sources per answer, being first stopped conferring much advantage, and the factor fell to 3.1% by Q3 2026. The second is the collapse of brand familiarity carried in model weights, from 24.1% to 4.8%. In early 2024, ChatGPT answered a large portion of commercial prompts from parametric memory alone, which strongly favored brands that were prominent in pretraining data. Almost every subsequent product change has pushed the assistant toward retrieving evidence at answer time, and the pretraining advantage has been priced out accordingly. The practical reading is that incumbency in the model's memory is now worth roughly one fifth of what it was worth two and a half years ago.
Recommendation Factor Weights by Purchase Category
The aggregate model conceals real differences between what ChatGPT rewards in enterprise software and what it rewards in a local trade. In the table below, we refit the model separately within six purchase categories, collapsing the four declining legacy factors into a single column.
| Purchase Category | List Mentions | Online Reviews | Awards & Accreditations | Customer Examples | Social Sentiment | Other Factors |
|---|---|---|---|---|---|---|
| B2B software | 41.7% | 12.4% | 9.8% | 18.3% | 11.2% | 6.6% |
| Professional services | 44.9% | 17.6% | 16.8% | 9.4% | 5.1% | 6.2% |
| Financial services | 40.2% | 15.3% | 20.6% | 8.7% | 5.9% | 9.3% |
| Healthcare providers | 33.8% | 24.1% | 22.7% | 7.9% | 4.3% | 7.2% |
| Consumer retail products | 29.3% | 31.6% | 4.2% | 6.8% | 19.4% | 8.7% |
| Local services | 26.4% | 38.9% | 11.3% | 6.1% | 8.8% | 8.5% |
- We found that list mentions lead in five of the six categories, with local services the lone exception, where review volume and recency carry 38.9% of the weight.
- Our data showed that awards and accreditations peak in healthcare at 22.7% and in financial services at 20.6%, the two categories where credentialing is a licensing matter rather than a marketing one.
- We found that customer examples matter most in B2B software at 18.3%, where named deployments and case studies function as the category's substitute for consumer reviews.
Recommendation Factor Weights by Prompt Type
The same brand can win one kind of commercial prompt and lose another, because ChatGPT weights the evidence differently depending on what it has been asked to do. In the table below, we split the prompt set into five types: open-ended discovery, meaning an unqualified request for the best option in a category; shortlist comparison, meaning a head-to-head between named brands; constrained recommendation, meaning a request bounded by geography, budget, or use case; agentic shopping research, meaning a multi-step buyer's guide compiled by the shopping tools; and follow-up validation, meaning a request to assess one named brand.
| Prompt Type | List Mentions | Online Reviews | Awards & Accreditations | Customer Examples | Social Sentiment | Other Factors |
|---|---|---|---|---|---|---|
| Open-ended discovery | 46.3% | 13.7% | 12.1% | 10.4% | 11.8% | 5.7% |
| Shortlist comparison | 24.8% | 21.3% | 13.6% | 20.7% | 12.4% | 7.2% |
| Constrained recommendation | 37.1% | 19.8% | 15.4% | 9.2% | 7.6% | 10.9% |
| Agentic shopping research | 21.6% | 34.2% | 8.3% | 7.4% | 13.9% | 14.6% |
| Follow-up validation | 18.4% | 27.9% | 10.6% | 12.8% | 24.1% | 6.2% |
- We found that open-ended discovery is the most list-driven prompt type at 46.3%, which follows from the assistant needing a defensible shortlist before it has any user constraints to work with.
- Our data showed that agentic shopping research inverts the aggregate model, putting online reviews first at 34.2% and dropping list mentions to 21.6%, because the shopping tools read product pages and rating aggregates directly.
- We found that social sentiment reaches its maximum on follow-up validation at 24.1%, where forum and community discussion of a named brand does more to shape the verdict than any other input.
Factor Weight Shifts Following Major ChatGPT Releases
Because our prompt set is held constant, each wave can be read against the product changes OpenAI shipped in the interval. In the table below, we pair nine dated releases with the factor whose weight moved most in the two quarters that followed.
| Date | ChatGPT Product Change | Factor Most Affected | Weight Movement (pp) |
|---|---|---|---|
| October 2024 | ChatGPT search launches with named in-line attribution | Search index position | +1.5 |
| April 2025 | Product results begin appearing inside standard answers | Online reviews | +2.4 |
| August 2025 | GPT-5 unifies the model lineup and routes queries to reasoning by default | Brand familiarity in weights | -4.5 |
| September 2025 | Instant Checkout and the Agentic Commerce Protocol launch | Online reviews | +1.9 |
| November 2025 | Shopping research launches on a shopping-tuned GPT-5 mini | Authoritative list mentions | +3.6 |
| March 2026 | In-chat checkout is deprioritized in favor of product discovery | Authoritative list mentions | +2.1 |
| April 2026 | GPT-5.5 ships with materially stronger browsing performance | Brand familiarity in weights | -2.8 |
| May 2026 | GPT-5.5 Instant becomes the default model for all users | Social sentiment | +1.7 |
| August 2026 | A dedicated shopping surface is pinned to the ChatGPT sidebar | Online reviews | +1.4 |
The pattern in this ledger is that model releases and retrieval releases move the weights in opposite directions. Every major model upgrade in the period, from GPT-5 in August 2025 through GPT-5.5 in April 2026, reduced the weight on brand familiarity carried in model weights, because each one shifted more of the commercial answer onto evidence gathered at answer time. Every retrieval or commerce release did the reverse for whichever external signal it exposed to the model: shopping research raised list mentions, the product surfaces raised reviews. We would note one limitation. Attribution to a single release is a judgment call when releases cluster, as they did in the autumn of 2025, and we report the largest mover rather than a decomposition. Readers wanting the full quarter-over-quarter matrix rather than the headline mover will find it in the dataset accompanying this report.
ChatGPT Releases Since This Wave
Our latest wave closed on August 28, 2026. On October 7, 2026, OpenAI made GPT-6 the default model in ChatGPT's Chat tab, with GPT-6 Sol for Plus, Pro, Business, and Enterprise users and GPT-6 Luna for Free and Go users starting October 8. The same release added Intelligent UI, which lets ChatGPT build an answer from text, visuals, and interactive elements, including side-by-side comparisons, and still give a plain text answer when that works best.
The weights in this report were measured before GPT-6 reached the Chat tab. In the ledger above, each earlier model upgrade lowered the weight on brand familiarity carried in model weights. We have not yet measured whether GPT-6 continues that pattern.
Sources
- Commercial Recommendation Benchmark Study, TGA Research Team, September 2026, New York, New York.
- Introducing shopping research in ChatGPT, OpenAI, November 24, 2025. https://openai.com/index/chatgpt-shopping-research/
- Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol, OpenAI, September 29, 2025. https://openai.com/index/buy-it-in-chatgpt/
- Introducing GPT-5.5, OpenAI, April 23, 2026. https://openai.com/index/introducing-gpt-5-5/
- GPT-6 and Intelligent UI for everyone, OpenAI, October 7, 2026. https://openai.com/index/gpt-6-for-everyone/
- ChatGPT Release Notes, OpenAI, October 2026. https://help.openai.com/en/articles/6825453-chatgpt-release-notes
- Model Release Notes, OpenAI, August 2026. https://help.openai.com/en/articles/9624314-model-release-notes
- Stripe powers Instant Checkout in ChatGPT and releases Agentic Commerce Protocol codeveloped with OpenAI, Stripe, September 29, 2025. https://stripe.com/newsroom/news/stripe-openai-instant-checkout
- Generative Engine Optimization (GEO): Explanation & Algorithm Breakdown, First Page Sage, 2026. https://firstpagesage.com/seo-blog/generative-engine-optimization-geo-explanation/
- Your LLM Has a Passport: Investigating Brand Origin Geographic Bias in E-Commerce Rankings, Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency, 2026. https://dl.acm.org/doi/10.1145/3805689.3806438