Which AI engines cite small business websites?

How was this measured?

First, one quick word. When an AI "cites" a website, it means the AI read that page and used it to build its answer. That is the thing we counted.

This study uses three Tenva checks from May and June 2026. Each one looked at a different kind of small business. The first check covered software companies. We asked 20 buyer questions, one time each, to ChatGPT, Claude, Gemini, and Perplexity. That gave us 80 answers with no errors. The second check covered 837 wedding venues in nine US metro areas. We compared each venue to the AI answers couples get when they look for a place to marry. On average, about 14.6 of those answers touched each venue. The third check looked at dentists in the Philadelphia area.

Two limits apply to every number here. We asked each question only once per engine, so think of these as snapshots. And AI answers change every time you ask. We only checked software companies, wedding venues, and dentists, so these patterns fit those markets, not every business. You can read the full method in the Tenva baseline. We run these checks every few months. The next one is set for Q3 2026.

How do the four engines differ in citation behavior?

For the software check, we wrote down three things about every answer. Did the engine use any websites at all? How many web pages did it use? And how many different websites did those pages come from? The four engines acted very differently. Perplexity and Claude used websites on all 20 of their answers. ChatGPT used websites on 19 of 20, and Gemini on 17 of 20. The bigger gap is in how many. Claude used 33 web pages per answer on average. ChatGPT used 5.3. So Claude reaches far more pages each time it answers.

Added up, all 80 answers used 1,182 web page links. The last column is empty for Gemini for a simple reason. Gemini wraps its sources in Google links, so we could not see which websites they really were this time. We plan to unwrap those links and find the real websites in our next run.

Here is what this means for a small business. The engine that gives you the most chances to show up may not be the one your customers use most. Claude used 659 web pages across just 20 answers. That means one answer can pull from dozens of pages, so a strong page has more ways to make the cut. ChatGPT used only 5.3 pages per answer, so the bar to get in is higher. Whether being used by an engine reaches real customers depends on who uses that engine. But these numbers show where the door is widest.

Who actually gets cited for small-business questions?

Leaving out the Gemini links we could not read, the most-used websites were a mix of small startup sites, marketing blogs, and a few bigger publishers. The top of the list is small startup tools, not famous media names. This matters for a small business. When the websites holding these spots are small, the spots are within reach. Your new page competes against a startup blog, not against a national outlet that has been around for decades.

There is a difference between getting named and getting cited. Getting named means the AI mentions your business in the text. Getting cited means the AI actually read your website to write the answer. One tool was named in 16 answers, but its own website was used in only 1. The two top tools were each used 27 and 23 times. A name shows the AI knows your reputation from somewhere else. A citation shows the AI read your page. The main business we tracked was never named and never used, across all 80 answers.

How often are small businesses named at all?

The wedding-venue check answers this on a big scale. We looked at 837 venues in nine metro areas. We compared each one to the AI answers couples actually get. Most venues never came up at all. The ones that did come up showed up only now and then. Getting named is the first step before getting cited. For most of these venues, the AI never even took that first step.

When a venue lost a spot, it usually lost it to a directory. For 59% of venues, the top rival the AI used was a wedding directory, not another venue. The site herecomestheguide.com was the top rival for 386 venues. The site weddingwire.com was the top rival for 106. The list below shows we covered well-known wedding markets across the country.

Other research points the same way. A 5W Public Relations report (PR Newswire, 2026) found that 73% of AI wedding-planning answers point to just two directory sites. It found that 84% of individual venues and vendors are used in zero answers. And it found that AI use among engaged couples doubled in a year, up to 36%. Put together, the Tenva check and the 5WPR numbers tell the same story from two sides. A handful of directories take most of the answer spots. Most individual venues are simply not there. And more and more couples are asking these questions. If a venue is one of the 68% that never gets named, it is not being punished. It just has not given the AI a page worth using yet. So the directories win by default.

Does Perplexity choose sources differently for local service queries?

The honest answer is yes, a little, but we should not overstate it. In the software check, Perplexity used websites on all 20 answers, with 7.8 web pages per answer. Those came from 115 different websites — the widest spread of any engine after Claude. So Perplexity reaches for many different websites, but it does not pack each answer with links the way Claude does.

The venue check shows what tends to win local spots, no matter which engine you use: clear, easy-to-compare facts. Directories took the top spot for most venues because they list facts an AI can trust and check, like how many guests fit, what the price tiers are, and where the place is. They list them the same way every time. That points to the one thing a single business can do: put the same kind of clear, checkable facts on your own pages. Keep in mind this is a look at two checks in two markets, not a strict test of how engines pick sources. So treat it as a pattern worth acting on, not a hard rule.

How often do AI assistants refresh their sources?

This study did not measure how fast engines swap sources. Each run is just a set of snapshots from one moment in time. One run cannot show how quickly an engine drops one website and picks up another. That is what the next run is for. By asking the same questions again in Q3 2026, we will see how the answers move over time.

What we can say is that answers change from run to run, even on the same engine. And the engines find websites in different ways. Perplexity and Gemini search the live web when you ask. ChatGPT mostly reads from a saved index instead of browsing fresh each time. That is why a business can show up in one check and not the next. It is also why the trend across many checks matters more than any single result.

So if you are deciding how often to check, remember this: one check is a reading, not a final grade. Missing once does not mean the AI will never use you. Showing up once does not mean you have won the spot for good. The real signal is the direction over time. A page that shows up in more answers across checks is gaining ground, no matter what one snapshot says. We check every few months. If you are actively adding new pages, check again two to four weeks after a change to see if the engines picked it up.

What should a small business do with these numbers?

Start by checking your own answers, not by guessing from these market numbers. Ask the questions your customers ask, across all four engines. Write down who gets named and which websites get used. The free checker hub does this for you. Then go after the open spots the data keeps showing — the 9 of 20 questions with no strong source, and the venue questions that directories dominate. You do that by publishing clear, fact-filled pages that each answer one buyer question. We explain how in how to get cited in AI answers and structured data for AI citations.

The dental checks show this works. A Philadelphia dentist who had done the work was used in 5 of 40 AI answers to patient questions. A similar dentist who had not done the work landed 0 of 40. A separate check found that 12 of 16 buyer questions had no strong source, across 55 of 64 engine calls. Those are open spots waiting for a good answer. Wedding venues can take the same path. The details are in AEO for wedding venues, and you can track your progress over time with an AI visibility score.