Why Does AI Recommend Different Businesses Every Time?

AI recommends different businesses every time because each answer is one draw from a bigger pool of businesses the engine could name, and the draw changes from one ask to the next. Ask Google AI Mode the same home-services question for the same city a week later and only about 4 in 10 of the businesses it named come back (41%); ChatGPT, the steadiest engine we measured, repeats 58%. Those numbers come from the Cheers Market Baseline, 34,742 answers from ChatGPT, Gemini, Perplexity and Google AI Mode collected between August 31 and September 20, 2026, and they mean one screenshot of ChatGPT tells you very little about whether customers in that city are hearing your name.

Last verified: September 23, 2026.

Every week, Cheers asks ChatGPT, Gemini, Perplexity and Google AI Mode the same 32 home-services questions in 100 metro areas across the United States and Canada, and records every business each engine names, with its star rating, review count and map position where the engine shows one, and every source it cites. Three weeks of that is a direct test of the question in the title: same question, same city, same engine, one week apart.

How much the list changes in a week

The chart at the top of this page is the core result. For every question and city, we compared the businesses an engine named one week with the businesses it named the next. Google AI Mode brought back 41.2% of them. Gemini brought back 42.6%, and ChatGPT 57.7%.

Identical lists were rare: 7.2% of ChatGPT's week pairs and 2.4% of Google AI Mode's. The other extreme was common. In 19% of Google AI Mode week pairs, none of the previous week's businesses came back at all; that happened in 5% of ChatGPT pairs and 21% of Gemini's.

Two caveats matter here, and I'd rather state them than let the chart oversell. We ask each question once per city per engine per week, so the change mixes whatever really moved (a new review, a new directory page, a model update) with the plain variation you would see asking the same question twice in a row, and three weekly checks can't pull those apart. Our name extraction also misses some businesses. A hand check of 40 answers found it caught about three in four of the businesses a reader would count, and 92% on Google AI Mode, so a missed name looks like a business that dropped out, which pushes the repeat shares down. Perplexity is left out of the comparison because a collection outage in the week of September 7 left it too few weekly pairs.

Why the same question gets a different list

Part of the churn is built into how language models write. They pick each next word from a spread of likely options instead of always taking the single most likely one, and OpenAI's API reference describes the setting that controls how wide that spread is:

Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic.

ChatGPT gives its users no way to change that setting, so typing the same words twice is no guarantee of getting the same answer.

Retrieval adds more movement. Google says AI Overviews and AI Mode may use a technique it calls query fan-out, "issuing multiple related searches across subtopics and data sources", to build a response, and its documentation is direct about the result:

AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary.

Google also says its generative AI features "are rooted in our core Search ranking and quality systems", so AI Mode is reading from the same kind of search results that decide the ordinary rankings. Each of those searches can return a slightly different set of pages, and a different page can mean a different business. The biggest reason in our data, though, is simpler than either mechanism: the pool is bigger than the list.

Google AI Mode typically shows about four business cards in an answer. Over three weekly checks it named a median of 8 different businesses for the same question and city, and so did ChatGPT. Most of those names were passing through. Only 12% of Google AI Mode's businesses were named in all three weeks and 64.5% appeared once; ChatGPT held on to more, with 25% named every week and 49.5% named once. Both sets of counts are floors, since a name the extraction missed can't be counted.

The slot almost never stays empty, either. Fewer than 1 in 500 ChatGPT, Google AI Mode or Gemini answers declined to name any local business. When someone asks who to call, the engines answer with somebody, so a business that isn't named was replaced rather than skipped.

Different engines keep different shortlists

If your business shows up in Google AI Mode and not in ChatGPT, you are looking at the normal case. When ChatGPT and Google AI Mode answered the same question for the same city in the same week, they shared about one business on average (1.04), and in 36% of those cases they shared none.

Across all four engines the split is wider. On the 5,347 occasions when ChatGPT, Gemini, Perplexity and Google AI Mode all answered the same question for the same city in the same week, 76.5% of the businesses named were named by only one engine, 16.2% by two, 5.8% by three and 1.6% by all four. Only 17% of those occasions had any business that all four engines named. Agreement is understated here too, because every missed name counts as disagreement, and Perplexity's names are the hardest for our extraction to catch. How the engines differ beyond the shortlist, from sources to booking links, is summarized in AI local ranking factors for 2026.

Why your business isn't showing up in ChatGPT

Rule out bad luck first. ChatGPT named a median of 8 different businesses for one question and city across three weeks, and about half of them appeared only once, so a single missing answer is weak evidence. Ask the same question on different days before deciding you are absent.

If you are missing across repeated checks, look at what ChatGPT reads. It leans on directories and "best of" lists far more than Google AI Mode does, and in our September 18 to 22 capture about a third of the rated business cards on its map (34%) linked to Yelp. The guide to getting ChatGPT to recommend your business breaks those sources down with the numbers, and why AI recommends your competitor shows how to compare your evidence with the business that won the answer.

Who keeps coming back

The businesses that stayed on the list tended to have deeper review histories and strong ratings. Among businesses Google AI Mode named at least once for a question and city, those showing 1,000 or more Google reviews came back in at least two of the three weeks 43.3% of the time. The share fell with each step down in review count: 41.1% for 200 to 999 reviews, 36.5% for 50 to 199, 30.4% for 20 to 49 and 25.7% for fewer than 20.

Rating mattered at the low end and not at the top. Among businesses with 100 or more reviews, 15.8% of those rated 4.9 were named in all three weeks, against 14.5% for a perfect 5.0, which is no real difference. Businesses rated 4.0 to 4.4 managed 6.7%. A 5.0 bought no extra consistency over a 4.9 once review counts were similar; a rating in the low 4s went with far less.

All of this is association. Review count also tracks how long a business has operated and how big it is, and we only see the businesses that were named, never the ones that weren't. For review counts by trade, see how many Google reviews you need for AI recommendations.

Consistent listings are the one input on this list we did not measure, so treat this paragraph as reasoning rather than a finding. An engine can only repeat a business it recognizes as the same business. When a location's name, phone number or address differs between its Business Profile, its website and the directories an engine reads, the engine is choosing between records that may not look like one company.

What this means for how you measure

A screenshot of ChatGPT is not a measurement. It is one draw, from one engine, in one place, on one day, and the next draw will almost never match it: only 7.2% of ChatGPT's week pairs repeated the same list. What an operator can measure is a rate: the share of checks, across repeated runs, engines and locations, in which the business is named. Run it the way a franchise tracks anything else that varies by market:

• Fix a small set of buyer questions for each location, with the city written into the question, and freeze the wording.
• Run every question on each engine your customers use, repeat it at least weekly, and wait for several runs before drawing a conclusion.
• Record who was named and which sources were cited, then report the share of checks that named you by engine and by location, never a single yes or no.
• Compare each location with the competitors that keep appearing in its answers, since those are the businesses taking your slots.

Location changes the answer too, and that needs its own test design; why AI search results change by location covers the controls. The multi-location audit guide shows how to organize the whole portfolio.

What to do this week

Pick your five highest-revenue locations and the three questions a customer would most likely ask about each. Run every question on ChatGPT and Google AI Mode twice this week, on different days, and write down each business named. By Friday you will have 60 answers, enough to sort locations into three piles: named in most checks, named sometimes, never named. The free AI Visibility Grader gives a one-business starting point across ChatGPT, Gemini and Perplexity if you would rather not start from a spreadsheet.

For the never-named pile, go to the reviews first. Count them, check the rating, and compare both with the businesses that keep taking the slot.

Methodology

Every week, Cheers asks ChatGPT, Gemini, Perplexity and Google AI Mode the same 32 home-services questions in 100 metro areas across the United States and Canada, and records every business each engine names, with its star rating, review count and map position where the engine shows one, and every source it cites. This analysis covers the three weeks from August 31 to September 20, 2026: 34,742 usable answers (9,434 from ChatGPT, 9,435 from Google AI Mode, 9,364 from Gemini and 6,509 from Perplexity) across 60 US and 40 Canadian metros. We excluded empty or cut-off answers, most of them from a Perplexity collection outage. Each question was asked once per city per week, so week-to-week changes include normal answer-to-answer variation. Business names were extracted automatically; a hand check of 40 answers found the extraction caught about three in four of the businesses a reader would count, so counts are lower bounds. Star ratings and review counts are the figures each engine displayed; for ChatGPT and Gemini we captured them from September 18 and September 14 onward. Findings describe associations in this sample, not how any engine ranks businesses.

For the repeat, pool and persistence figures we compared consecutive weeks for the same question, metro and engine. ChatGPT and Gemini businesses are the ones named in the answer text, because their map and place cards were captured only late in the window; Google AI Mode includes its place cards, which we captured in all three weeks. Businesses were matched across weeks and engines by normalized name within each metro. The pool figures cover the 2,861 Google AI Mode and 2,799 ChatGPT question and city pairs answered in all three weeks, which named at least 23,949 and 23,507 business, question and city combinations; the four-engine overlap covers 5,347 occasions. Review bands use the highest review count Google AI Mode showed for a business in that metro. Perplexity is excluded from the repeat, pool and persistence figures because of its outage.

Sources

Checked September 23, 2026.

• Create a model response. OpenAI API reference, checked September 23, 2026. Source of the quoted description of the temperature setting.
• AI features and your website. Google Search Central, last updated December 10, 2025. Source of the query fan-out description and the quoted statement that AI Mode and AI Overviews responses and links vary.
• Optimizing your website for generative AI features on Google Search. Google Search Central, last updated July 10, 2026. Source of the quoted line that Google's generative AI features are rooted in its core Search ranking and quality systems.
• Cheers Market Baseline, August 31 to September 20, 2026: 34,742 answers from ChatGPT, Gemini, Perplexity and Google AI Mode to 32 home-services questions in 100 US and Canadian metros, aggregate only. Research methodology. Source of every repeat, pool, overlap, review-band and rating figure in this article.

Dylan Allen-Arnegård is the CEO & Co-Founder of Cheers, the local search platform for multi-location service businesses.