What AI assistants cite when owners ask about AI receptionists

From September 9 to September 25, 2026, we put the same 19 questions an owner asks before buying an AI receptionist to two AI assistants with web search, 13 times each, and kept every answer. This page reports which companies the answers named, which pages they cited, how much that changed between runs and where the cited evidence is thin. Chleb AI sells an AI receptionist, so it also reports how often the answers named us.

The short version

Across 424 answers to buyer questions:

  • Claude named at least one company in 206 of its 247 answers. Gemini did in 114 of its 177.
  • Most cited pages sat on the website of a business that sells something: 1,871 of Claude's 1,932 citations (97 percent) and 777 of Gemini's 898 (87 percent).
  • Claude's sources were steady from run to run, with an average overlap of 0.70 between two answers to the same question. Gemini's were not, at 0.23.
  • For 16 of the 19 questions, no cited page on an independent site claimed, in its address or title, to have tested the products.
  • Chleb AI was named or cited in 2 of the 424 answers, both from Gemini, for the question "AI receptionist company in Tampa Florida".

How the answers were collected

The answers come from Chleb's own nightly measurement job, which asks each question through the command-line versions of two assistants. Claude runs as claude-sonnet-5 with only its web search and page reading tools switched on. Gemini runs as gemini-3.1-pro-low through the Antigravity command-line tool with its web search tool. Each question starts in a new, empty working folder, Claude with no saved settings or session history. Both get the same question prompt: a US small business owner asks this; search the public web for the exact question first, answer in under 180 words, cite sources and list every source address used. Claude also gets a short system prompt that asks it to act as a neutral public-web research assistant, cite the sources it actually consulted and not favor a brand. Gemini gets no system prompt from the job. Both run with the tools' effort setting at low.

There are 19 buyer questions, listed in full in the explorer and the data file: what an AI receptionist is and what it costs, how it compares with an answering service or a hired receptionist, booking and missed-call text back, one question each for HVAC, plumbing, dental, salon, med spa, auto dealer and lawn care businesses, alternatives to Smith.ai and Goodcall, CRM setup, a website with booking, AI receptionist companies in Tampa, the best AI receptionist for a small business, and an emergency plumber answering service in Tampa. A further question asks about Chleb AI by name and is reported on its own.

A run counts when it recorded every question for an engine with the same prompt and a verified exit location. 13 runs qualify, from September 9 to September 25, 2026, all from a connection that exited in the United States. Claude answered all 247 buyer questions it was asked. Gemini answered 177 of 247: its usage quota stopped 62 of the other 70, and on 3 of the 13 runs it answered nothing. A blocked question is missing data, not a negative result.

On September 21, 2026 the job ran from a connection that exited in Ireland. That run is kept out of the main figures, because answers from different places are not comparable. Its 4 back-to-back Claude passes, started between 19:10 and 19:19 UTC, are reported as their own series to show how much answers move within minutes. The pass started at 19:19 UTC stopped after 13 of the 20 questions.

  • Left out, September 5: one run, because it saved a summary of each answer in an earlier format, not the answers and cited pages this analysis reads.
  • Left out, September 5 to 25: 35 run folders, because they saved no answers (failed, skipped or test runs).
  • Left out, September 9: one run, because it used an earlier wording of the prompt and did not verify its exit location.
  • Left out, September 9: one run, because it asked 2 of 20 questions.
  • Left out, September 21: a Claude retry pass, because it answered 7 questions, fewer than 10.
  • Left out, September 21: a Gemini retry pass, because it answered 0 questions, fewer than 10.
  • Left out, September 21: the nightly pass, because it exited in Ireland, so it is not comparable with runs from the United States; only that evening's repeat passes are reported, as their own series.

Who gets named

A company counts as named when the answer's own text names it, using a reviewed list of 187 company and product names. A company that appears in the question does not count, such as Smith.ai in a question about Smith.ai alternatives. Neither does a link whose visible text is only the name of the site it points to, because that labels a source rather than suggesting anyone.

Claude named at least one company in 206 of 247 answers (83 percent), 5.4 per answer on average and 123 different companies in all. Gemini named at least one in 114 of its 177 answers (64 percent), 2.8 per answer. The companies Claude named most were Aira, Rosie and Smith.ai; Gemini's were Smith.ai, Rosie and Goodcall.

Some questions drew explanations instead of names. Claude named a company in at most one of its 13 answers to "what is an AI receptionist and how does it work for a small business", "AI receptionist vs answering service vs hiring a receptionist" and "GoHighLevel CRM setup and calendar booking for a service business".

One question was read two ways. For "best emergency plumber answering service in Tampa", Claude's answers cited plumbing companies' own websites 84 times over 13 runs, as if a homeowner needed a plumber. Gemini's 8 answers cited no plumbing company sites and named Answer Excellence and AnswerFirst most often, which the answers describe as answering services.

Chleb AI was named or cited in 0 of Claude's 247 buyer answers and 2 of Gemini's 177. Asked about "Chleb AI receptionist" by name, Claude cited chleb.ai in 1 of 13 answers and named other companies in 11 of them; Gemini cited chleb.ai in 5 of its 8.

What gets cited

Claude cited 1,932 pages across its 247 answers, on 152 websites. Gemini cited 898 pages on 274 websites, and 60 of its 177 answers listed no source pages at all.

We classified the owner of every cited website by hand: well-known platforms, directories and publications by name, the companies on our vendor list as sellers, and every other site against its own homepage or the cited page. A cited page counts as company-owned when it sits on the site of a business that sells products or services, whether software, phone and answering services, marketing or trade work. It counts as independent when it sits on a directory or review site, a forum, a news, trade or independent publication, or the site of an owner that sells nothing to the businesses asking, such as an investor. Videos, social media and user-published posts, such as Medium articles, do not show their author and are counted apart. By that measure 1,871 of Claude's citations (97 percent) and 777 of Gemini's (87 percent) were company-owned: product pages, blog posts, press releases and help pages. Independent sites made up 60 of Claude's citations and 81 of Gemini's. For 4 of Gemini's citations we could not establish the owner, and they are counted in neither group.

Best-of lists and comparisons are a large share. Claude cited 670 of them, and 619 of those sat on the website of a company in our vendor list, a seller ranking the field it competes in. Gemini cited 250, 178 of them on vendor sites.

The two engines lean on different kinds of sources. Gemini cited YouTube in 27 answers and Reddit in 21. Claude cited no videos and no forum threads in its 247 answers.

Claude's citations follow its search closely. It ran a single web search in 242 of its 247 answers. Of its 1,932 citations, 1,930 were results of those searches, and of the 2,123 results the searches returned, it cited 1,930. In practice, Claude cited nearly every result its searches returned and almost nothing else. Gemini's command line does not show its search results, so the same check is not possible for it.

How much the answers change

To measure change we compared the websites cited in every pair of answers to the same question, using Jaccard overlap: the websites both answers cited, divided by the websites either one cited. Two answers citing the same sites score 1, and two with nothing in common score 0. Answers that listed no sources are left out.

Claude was steady. The average overlap across its 19 questions was 0.70, and 0.74 between answers from consecutive runs, so answers close in time were a little more alike than answers weeks apart. For 18 of the 19 questions, at least two websites appeared in every Claude answer.

Gemini was not. Its average overlap was 0.23 by website and 0.18 by page, and for 13 of the 19 questions no website appeared in every Gemini answer that listed sources.

Asking again within minutes does not settle it either. In the 4 Claude passes run back to back on September 21 from Ireland, the average overlap was 0.60. An answer sampled once is a weak basis for any claim about what AI assistants recommend.

Where the cited evidence is thin

We read each cited page's address and its title for three kinds of proof a buyer might want: a test of the products, a price list from the seller, and local facts for a local question. Titles come from Claude's search results. Gemini shows none, so a page Gemini cited has a title here only when a Claude search in the data also returned that page. The pages were not opened for these checks, so this is what the citations say about themselves.

  • Tests: 30 of Claude's citations and 17 of Gemini's were pages whose address or title says the author tested, tried or compared the products. Of those, 13 sat on independent sites, all on two websites: smash.vc and reddit.com. For 16 of the 19 questions, no independent page claiming a test was cited.
  • Prices: Claude quoted a dollar amount in 143 of 247 answers and Gemini in 29 of 177. Vendor price lists were cited for one question, "how much does an AI receptionist cost per month": 23 times by Claude and 4 by Gemini. For 14 questions, answers quoted prices while citing no price list.
  • Local facts: for "AI receptionist company in Tampa Florida", no cited page about the area sat on an independent site such as a directory, review site or local publication.
  • Independence: for all 19 questions, at least eight in ten cited pages were company-owned.

What this means if you are choosing a receptionist

Most of the rankings these answers cite were published by companies that sell in the field they rank. The three things these answers rarely supply are the ones you can collect yourself: a test call to each receptionist on your shortlist, the seller's own price list with usage charges spelled out, and a reference from a business near you.

Chleb's prices are on the pricing page, and you can hear Chleb's own AI receptionist by calling (813) 694-8671.

What this does not show

Read the figures above with these limits in mind.

  • Consumer apps: these are command-line versions of Claude and Gemini with a fixed prompt, a system prompt for Claude and the effort setting at low, not the Claude, ChatGPT or Gemini apps, Google AI Overviews or Perplexity, which add their own ranking, personalization and location.
  • Quality: being named is not being good. Nothing here tests the products or checks what the answers say about them.
  • The pages themselves: page types and evidence signals come from each page's address and, where a Claude search returned it, its title. A page may say more than its title, and may have changed since it was cited. Site owners were checked once, on September 28, 2026.
  • Every company: names outside our list of 187 are not counted, and names that are also ordinary words, such as Ruby or Weave, only count with a capital letter.
  • Other times and places: 13 runs over 17 days from one exit country. The repeat passes from Ireland show that answers can differ by place and by minute.
  • Gemini's gaps: 70 of its 247 buyer questions were blocked and 60 of its answers listed no sources, so its figures rest on fewer answers than Claude's.
  • Who published this: Chleb AI sells an AI receptionist and is one of the companies these answers could name. The data file has every figure on this page so you can check them.

Questions buyers ask

Does being named by an AI assistant mean a product is good?

No. The answers here mostly repeat pages that businesses published about their own products or their own field. Across all 424 answers, the cited pages on independent sites whose address or title claims a test or comparison came from two websites: smash.vc and reddit.com. Treat a name in an AI answer as a lead to check, not a verdict.

Why do Claude and Gemini answer the same question so differently?

They find sources differently. Claude's answers here cited nearly every result its web search returned and little else. Gemini's cited sources changed from run to run, and 60 of its 177 answers listed no sources at all.

Can I download the data?

Yes. The CSV on this page has one row per run series, question and assistant, plus an all-questions total for each series and assistant, with the counts, the most named companies, the most cited websites, the overlap figures and the evidence flags. It contains no answer text and no page addresses.

Sources

  1. The data behind this page: one row per run series, question and assistant, with totals (CSV), read 2026-09-28

Book a demo