Who AI recommends, and why it changes: the dated record, October 2026

1. Only Perplexity is sticky

Perplexity API answers named the same business again the next answered day in 1,750 of 2,795 pairs (0.6261), against 137 of 681 for OpenAI (0.2012) and 286 of 1,000 for Gemini (0.2860), 2026-09-01 to 2026-09-27.

What this means for a local business

A business that Perplexity names today is named again the next answered day about 63 in 100 times in this record, against about 20 in 100 for OpenAI. The same business showing up in one check is therefore a stronger signal on Perplexity than on the other engines. This is a description of past answers, not a prediction for any business.

Limits

• Window is 2026-09-01 to 2026-09-27 for OpenAI, Gemini and Anthropic. Perplexity collection changed in the files dated 2026-09-27, so the Perplexity cited-domain figures in finding 3 use 2026-09-01 to 2026-09-26. Perplexity figures in findings 1 and 4 and the listing comparison were computed through 2026-09-27 and have not been recomputed for the 2026-09-26 cut. Anthropic persistence rests on 38 pairs and churn on 44, so it is not compared.
• Names are keyed per market after alias merging, so identity across days relies on the merged key.
• API answers only; consumer apps can behave differently (see finding 2).

2. The door changes the answer

Asked the same 9 Pleasanton questions, the logged-out ChatGPT app named local businesses on 8 of 8 answered questions (2026-10-05), while the OpenAI API with web search off named a Pleasanton local firm on 0 of 9 (2026-10-04).

Prior work: info.link, ChatGPT API vs UI: do they return the same citations? (2026-08-31).

What this means for a local business

Checking how an AI assistant answers about your business through one door does not tell you what customers see through another. In this test the logged-out app answered with named local businesses and the API call, which ran without web search, answered with a checklist plus directories and chains. We disclose the web-search setting because it is the mechanism that differs. This extends prior work: info.link published a ChatGPT API vs UI study on 2026-08-31 (52,170 chats, 41% brand-visibility difference, brands generally). Our addition is local businesses, with the web-search setting disclosed.

Limits

• The days differ: API on 2026-10-04, app on 2026-10-05.
• Which OpenAI model the API row used is not recorded in the file.
• The cloud browser's location is unknown; the questions name the city.
• The business names in the app column are an extraction by our lane. Only the question 5 answer was confirmed against screenshots, where the logged-out app rendered a local map of Dublin and Pleasanton with rated place pins.
• Question 9 on the app side ended in an error, so the app denominator is 8, not 9. Logged-out Perplexity and Copilot hit login walls and are not compared.

3. Directories feed the answers

Perplexity cited expertise.com on 248 of 642 cell-days with cited sources (38.6%) and yelp.com on 214 of 642 (33.3%), 2026-09-01 to 2026-09-26.

What this means for a local business

The pages Perplexity cites most often for local questions in this record are directory and review sites. Whether a business is listed on them is a fact you can check. Among businesses Perplexity named at least once, those on the Expertise.com page were named on a median of 5 days against about 1 day for those not on it. This is an association; most listed businesses were not named at all in the window.

Limits

• The Expertise.com result is a correlation, not a cause. Listed firms tend to be established and review-rich, which can independently drive naming.
• Listing membership was read from live pages on 2026-10-05 and may differ from September membership.
• Yelp could not be scraped for the listing comparison, so no Yelp comparison was made.
• The comparison covers 15 cells; 14 cells had no Expertise.com page cited. Pleasanton injury-law cells may be affected by other work and show the largest effect.
• Domain classification is heuristic. Google SERP rows are partly missing: 791 of 1,269 SERP rows failed extraction at a consent wall and carry no data.

4. The engines rarely agree

The engines rarely named the same businesses: the highest pairwise agreement was Gemini with Perplexity at 0.0610 (17.0723 of 280 shared answered cell-days), and Perplexity with Anthropic scored 0.0000 (0 of 855), 2026-09-01 to 2026-09-27.

What this means for a local business

Being named by one engine tells you almost nothing about being named by another, so each engine is a separate check. Gemini spread its names across the most distinct businesses and Perplexity across fewer, so the share of a market that any one business can hold differs by engine.

Limits

• Anthropic named a local business on 34 of 1,026 answered cell-days, so its rows are thin and agreement with it is close to zero by construction.
• Pooled effective-business figures are a mean of cell values, not a business-weighted figure.
• 108 cell-day-engine duplicates were merged by union, and the 2026-10-04 Google SERP quarantine list was applied.

5. Most popular AI-visibility statistics do not survive a trip to their source

Of 22 popular AI-visibility statistics traced to their source on 2026-10-05, 5 were sourced, 7 distorted, 8 unsourced, 1 circular and 1 contradicted.

Two examples from the ledger: the statement that 88% of local businesses are invisible in AI search was graded unsourced, because the page that carries it names no study. The finding that 83% of restaurants were invisible in ChatGPT results was graded sourced, because a primary study states it, run by a vendor on restaurant queries only.

What this means for a local business

Before relying on an AI-visibility statistic in a sales pitch or a budget, trace it to the page that first states it, its date, its sample and what was actually measured. Each graded statement is listed with its sources on the claims page.

Limits

• Sourced means the number exists at the primary source and says what is claimed. It does not mean the method was checked.
• Some pages could not be scraped (social posts and forum threads were seen only as search excerpts), so those statements are not in the 22.
• Grades are our reading on 2026-10-05; the claims page lists each source so you can check it.

Method

• The dated record started on 2026-08-02 and covers 27 markets. Fixed questions for each market and niche are run every day through the engines.
• Window used for the figures above: 2026-09-01 to 2026-09-27, API answers, 5,076 rows, 38 market-and-niche cells, 4 engines. Perplexity changed its collection method in the files dated 2026-09-27; see Limits for which figures use 2026-09-26 as the last Perplexity day.
• Google rows listed in the 2026-10-04 quarantine file list are excluded.
• Raw files are hashed so a later change to any row is detectable.
• Definitions: persistence, churn, concentration, agreement and name rate are defined in the metrics file; an answered day is a day with at least one business string returned.

How to cite

Namebeam dated record, markets.namebeam.ai/report/october-2026/, published 2026-10-05