Houston AI Citation Study
October 2026.

• AI Mode cites few sources. It averaged 3.4 citations per answer against 10 for Perplexity, and most AI Mode citations on recommendation questions were the Business Profile cards of the businesses it named.
• Perplexity backs each pick with a source. A business's own website was among the sources for 38% of the businesses Perplexity recommended. The other 60% rested only on third-party pages such as Yelp, Zocdoc, Best Lawyers or HAR.com.
• Reddit barely appears. It was cited in 9 of 464 answers. For local service questions, both engines went to businesses, directories and Google's own listings instead.
• City names get confused. 9 of 58 answers about Pasadena, Texas cited businesses in other states, mostly Pasadena, California.

• The order of events holds. We also looked only at the 1,580 census businesses in these industries, using checks measured on October 2, before any question was asked. 5.9% of sites passing all four basics were cited by October 30, against 1.4% of the rest (adjusted odds ratio 4.6, 95% interval 2.3 to 9.2). FAQ schema (7.4% against 1.9%) and question-style subheadings (5.6% against 1.4%) showed the same pattern.
• Links do not explain it. Cited sites were not more linked-to than the census: their median Common Crawl rank was 18.2 million, against 14.5 million for the census (a higher number means fewer links). The gap held in every authority quarter, from 81% against 32% among the most-linked sites to 55% against 26% among the least-linked.
• We then tested 40 factors. Against Houston sites in the same industries, cited sites still passed all four basics more often (68% against 41%, odds ratio 3.0). Of 40 commonly recommended factors, six held up in every test: reviews on the homepage, a specific business type in the schema, prices, years in business, city pages and a meta description. See which AI citation factors matter.
• This is association, not proof. Businesses that fix the basics often invest in their site and reviews in other ways too. A controlled test by Ahrefs found that adding schema to pages that were already cited did not raise citations; our data covers a different case, local sites that AI had not necessarily seen before.

• City and location pages were the most-cited business pages (40%), ahead of homepages (21%), service pages (16%) and pricing pages (14%).
• For "who is the best" questions, 80% of cited business pages named the city in the URL or title, and 61% named it in the H1.
• For price questions, pricing pages carried the business. 41% of the business pages AI Mode cited on cost questions were pricing or cost pages. Two in three pages cited for price questions had FAQ schema.
• Cited pages were substantial. The median cited page had 1,235 words readable without JavaScript; 73% had business schema, 57% had FAQ schema and 67% used question-style subheadings.

• Self-ranking posts were rare: 4.2% of the business pages Perplexity cited and almost none in AI Mode.
• They gave no measurable edge. When a business's own "best of" post was cited, Perplexity recommended that business 47% of the time, about the same as for businesses whose other pages were cited (45%).
• Third-party ranking sites matter more. "Best ___ in [city]" lists on directories and ranking sites made up 20% of all Perplexity citations. Being listed on them is a better use of effort than writing your own.

• The sources barely moved. The cited domains overlapped 98.5% between the two runs.
• The top pick usually held: the same business came first in 82% of answers.
• The full list shifted. Only 69% of the recommended businesses overlapped between runs, so a business on the edge of a list can appear one minute and drop out the next. One check is not a measurement; track a set of questions over time on more than one engine.

ChatGPT and Claude here are their web search APIs on the cheapest models (gpt-5-nano and Claude Haiku 5.5), asked on October 31; AI Mode and Perplexity were asked on October 30. Each engine attaches sources differently: AI Mode through its reference list, Perplexity through numbered citations, ChatGPT and Claude through links inside each list item. "Same top pick when asked again" compares two runs about 6 to 10 minutes apart.

• Directories carry ChatGPT and Claude. On "who is the best" questions, 68% of ChatGPT's citations and 65% of Claude's went to directories, review and ranking sites. Business websites got 23% and 29%. Only 14 to 15% of the businesses they named had their own website among the sources for that pick, against 38% on Perplexity.
• Claude reads Yelp; ChatGPT never did. Yelp appeared in 88 of Claude's 120 "best" answers and made up 20% of all its citations. ChatGPT cited Yelp zero times; its most-cited sources were BestProsInTown (69 citations) and Birdeye (48). Which directories matter depends on the engine.
• Ranking pages feed Claude. 67% of Claude's citations on "best" questions were "best of" or "top 10" pages, against 29% for ChatGPT.
• Four engines, four different lists. Across the 120 recommendation questions the engines named 1,582 different business and question pairs; 79% were named by only one engine and just 24 by all four. Businesses are matched by name, so one written two different ways can be missed; the true overlap may be a little higher. See what the 24 had in common. ChatGPT's top pick matched AI Mode's 15% of the time and Perplexity's 12%; Claude's matched 7.5% and 9.2%.
• Consensus picks have big review counts. The businesses that both ChatGPT and AI Mode named had a median of 589 Google reviews, against 323 for all of AI Mode's picks.
• Price questions still go to business sites. On cost questions, business websites took 66% of ChatGPT's citations and 64% of Claude's, against 57% for Perplexity and 36% for AI Mode (plus 22% to booking pages). Every engine leaned on business sites for prices.
• ChatGPT is the least consistent. Asked again a few minutes later, ChatGPT kept the same top pick 36% of the time and Claude 56%, against 82% for Perplexity. 17% of ChatGPT's answers had no linked source at all.

See which SEO agencies the four engines named most: Houston SEO agencies AI recommends.

• Questions: 15 industries in 8 areas (Houston, Katy, Sugar Land, Spring, The Woodlands, Pasadena, Clear Lake and Humble). Each industry had a "Who is the best ___ in [city], TX?" question and, except restaurants, a price question such as "How much does AC repair cost in Katy, TX?". 232 questions in all.
• Engines: Google AI Mode, collected through SerpApi with the search located in Houston, and Perplexity Sonar through its public API. Each question was sent alone, with no system prompt and no brand named. All answers were collected on October 30, 2026 between 12:46 and 13:17 UTC; Perplexity's second run followed a median of 10 minutes later. ChatGPT and Claude were added on October 31, 2026 (12:34 to 13:01 UTC): every question twice, through the OpenAI Responses API web search tool (gpt-5-nano) and the Anthropic web search tool (Claude Haiku 5.5), the cheapest models each company offers with web search, located in Houston, with at most two searches per answer and no system prompt. The 928 answers cost about $11 in API fees.
• Citations: every source URL each answer cited, counted once per answer: 3,102 in the main run (5,422 including Perplexity's second run) from 1,034 domains. Each domain was coded by hand or by rule into a source type; local business sites were confirmed through the Index census, a Texas address or phone in their schema, or a match to a recommended business. Business sites with a non-Texas address were coded as outside Houston.
• Recommendations: AI Mode's Business Profile cards and named list items, and the bold business names in Perplexity's answers, with the sources each answer attached to that name. A hand check of 55 random Perplexity recommendations found 53 correct names; the two name errors and four website-matching errors it found were fixed before the final analysis. For ChatGPT and Claude, recommendations are the businesses named in list items, with the links inside each item as its sources; hand checks of 50 random recommendations each found 48 correct names for ChatGPT and 48 for Claude (the errors were a firm name split in two and a name cut short). The 396 domains they cited that the first round had not seen were coded by hand the same way.
• Website checks: cited local business homepages and every cited business page were crawled on October 30, 2026 with the same engine as the Index (October 2, 2026). Link authority is the domain's harmonic-centrality rank in the Common Crawl web graph for August to October 2026. Odds ratios come from logistic regressions with industry fixed effects; percentage intervals use a bootstrap over answers.
• Limits: one metro, two days (AI Mode and Perplexity on October 30, ChatGPT and Claude on October 31), and API or scraped results rather than a logged-in person's screen. AI Mode's citations come from its reference list, which may not show every source it used. The census covers businesses listed with a website in OpenStreetMap, not every business. Google AI Overviews were not included. ChatGPT and Claude were measured with their cheapest search models through the API, which can answer differently from the default models in the consumer apps, and on the next day.
• Disclosure: Marketing & AI is a Houston SEO and AEO agency, and SEO agencies were one of the 15 industries. We were not cited or named in any answer.