I asked ChatGPT, Perplexity & Google AI which restaurant website platform to use. Here’s what they said.

The experiment — and why I ran it

The question I get every month from operators — in the host stand at Tacombi, in The Window inbox, at the end of every audit call — is some version of the same thing. What should we actually use for our website? Wix? Squarespace? WordPress? Something a developer builds from scratch? A POS-bundled site from Toast or Square?

For the last twelve months that question has been changing under us. Operators are not just typing it into Google anymore. They are asking ChatGPT, Perplexity, and Google’s own AI answer the same question, getting three names back in a single paragraph, and treating the response like a recommendation from a friend. Whether that answer is good or thin, the diner-side equivalent is happening at the same time: prospective guests are asking AI assistants where to eat, and the assistants are quoting the same kinds of sources back about your restaurant. The two questions are the same shape.

So I ran the experiment I would want to read if I were one of those operators. On the morning of May 30, 2026, I asked the three AI engines the same three prompts about restaurant website platforms, in one sitting, in three clean private windows. I logged the recommendations, what each engine surfaced (or didn’t) about citation sources, and where each engine’s answer was either useful or honestly a little thin. This is a snapshot of one morning in late May 2026; the engines change their weighting often, and the same prompts a quarter from now will land somewhere slightly different. That is the point of stamping a date on a dispatch like this.

The three prompts I ran were these — deliberately operator-shaped, not researcher-shaped, because the goal was to read the engines the way a working restaurant owner would actually ask the question:

Three prompts, three engines, nine responses. Every one of them logged, screenshotted, and dated. The transcripts and the citation logs are referenced throughout the next four sections; the patterns are what an operator can actually use.

What the three engines recommended

The first thing to read is not which platform “won.” The engines do not vote. The first thing to read is the shape of the recommendations — how many platforms each engine named across the three prompts, where they agreed, and where each one reached somewhere the others did not. The shape is the clue about how each engine assembled the answer.

What I expected before I ran the prompts: a tight overlap on three or four well-known names — the platforms that appear in every “best of” roundup — and divergence on the long tail, where each engine reaches into a different pool of secondary sources. That is roughly what happened. Five platforms landed in all three engines: Squarespace, Wix, Toast, Square, and ChowNow. Four more landed in two of three. Four singletons each appeared in exactly one engine, and those four are the most interesting data point in the experiment, because each one is a small confession about whose framing that engine has absorbed about restaurant tech.

The five consensus platforms map cleanly onto how an operator actually picks one. Squarespace and Wix are the “I want a website that looks nice and works on a phone” answer. Toast and Square are the “I already have a POS and want the website to follow” answer. ChowNow is the “I want direct ordering without the third-party app” answer. Five platforms, five recognizable operator situations — this is exactly the kind of overlap you would expect from engines that have all read roughly the same restaurant-industry coverage over the last two years.

The 2-of-3 tier is where the engines start showing their hand. Three platforms — BentoBox, Owner.com, and GloriaFood — were named by both ChatGPT and Google AI but never by Perplexity. That is not random. ChatGPT and Google AI are reaching into the restaurant-specific source pool: trade roundups, vendor comparison posts, the kind of content that runs on industry blogs. Perplexity is reading somewhere else. WordPress — named by ChatGPT and Perplexity, missed by Google AI — is the inverse pattern: Google AI appears to be deprioritizing self-hosted answers for this question, while ChatGPT and Perplexity still treat “use WordPress” as a legitimate operator path.

The four singletons each say something about the engine that named them. ChatGPT was the only engine to surface SpotOn Restaurant, which has the same shape as the others on the commission-free list but appears in fewer industry roundups; ChatGPT reaches deeper into the long tail of named brands than the other two. Perplexity was the only engine to name Shopify as a restaurant-website option, framed correctly as “only if your restaurant is really acting like a retail business too”; that recommendation reads as a Shopify-marketing source seeping in. Perplexity was also the only engine to surface RestaurantDirect by EatStreet, which is a real platform but a much smaller name — another sign of Perplexity reading a different source pool. Google AI was the only engine to name Canva Website Builder for restaurants, which is the most genuinely off-base call of the experiment: Canva is a graphic-design tool with a website builder bolted on, and recommending it for a restaurant suggests Google AI lifted from a generic-business roundup rather than a restaurant-specific one.

For the evergreen platform-by-platform comparison that this dispatch links up to — the one I use on every audit call as a calmer, longer reference — see the best restaurant website platform comparison. This dispatch is the dated read; that piece is the slow-moving one.

Who they cited — and what that tells you about how AI search picks restaurant content

Most operators read an AI answer for the recommendation. The more useful read is the citation log underneath it. The recommendation tells you what the engine concluded; the citations tell you why. The engines cannot verify a website platform’s actual fit for a restaurant the way they can verify, say, an opening time off your Google Business Profile — they are summarizing third-party prose. So they lean on the sources that prose came from, and which sources they trust is the part operators can actually act on.

The citation log from the experiment is the second piece of data this dispatch turns on, and it is also where the experiment ran into its sharpest methodological limit. Each engine handles citations differently, and what an operator can lift back out of the answer depends entirely on which engine they asked.

ChatGPT, in default signed-out mode, surfaced zero citation links. The answers came back as structured prose — numbered lists, recommendation tables, and a “tell me more and I’ll narrow it down” closer — with no link, footnote, or source attribution anywhere. The platforms it named were pulled from training data, and the operator has no way to check the underlying source. That is not a bug; it is the shape of ChatGPT’s default mode. If you want citations from ChatGPT, you have to explicitly turn on web search.

Perplexity normally shows numbered citations underneath each claim — that is the whole architecture of the product. In this run, the citations were visible in the Perplexity UI but did not transfer cleanly when I copied the response into my log; only the prose came along. That is a capture limitation, not an engine limitation, and a future re-run of this experiment will capture the URL list properly. The recommendations Perplexity made are still legible — Shopify, ChowNow, Square Online, Toast, RestaurantDirect — but the verifiable trail back to which sources it leaned on is not in this log.

Google’s AI Overview included inline numbered citation markers (you can see them as [1, 2, 3] in the live answer), and the answer references specific facts that suggest specific source types: a $2.6% + $0.10 processing-fee figure for Square that reads like a Square pricing-page lift, a $119/month ChowNow figure that reads like a ChowNow comparison-page lift. The source URLs themselves were not captured in this run.

The honest read of the citation question, then, is twofold. Methodologically: only one of the three engines makes citations easy to capture by default, and even that one requires the operator to copy a separate list. Substantively: what we can infer from the answer text itself is that the engines are reaching for vendor pricing pages and industry comparison posts for the specifics, and for the general “best platform” recommendations they are reaching for the same well-known roundups every operator has already seen. A re-run of this experiment with proper citation capture is on the editorial queue.

The operator lesson, even with the citation gap, is the same: the engines are reading the same kinds of sources to recommend a platform that they read to recommend a restaurant. If your restaurant’s most-quotable answer to “is it good for a group of eight” or “do you have a vegan entrée” lives on a vendor-style page, on a roundup, on a community forum, or nowhere — that is roughly the order in which an engine will pick it up. The library piece on how to get cited in Google’s AI Overviews walks the mechanism in full; this dispatch is the field-test that confirms the same mechanism applies to platform-recommendation questions too.

Where the answers were wrong or thin

Here is the surprise of the experiment. I expected to write this section about engines parroting vendor “commission-free” copy without the fine print underneath. That is not what happened. On the commission-free prompt, all three engines named the processing-fee caveat themselves, in their own words, unprompted.

ChatGPT wrote: “Even ‘commission-free’ systems usually still charge: Credit card processing fees (typically around 2–4%), Monthly software subscriptions, Setup fees in some cases. The real distinction is that they don’t take a percentage of each order as a marketplace commission.” That is exactly the framing I would want an operator to walk away with.

Perplexity wrote: “‘Commission-free’ usually means no platform cut on orders, but you may still pay normal card-processing fees and sometimes monthly software fees. That distinction matters because it affects your real cost per order.” Same posture — the engine flagged the caveat and named why it mattered.

Google’s AI Overview was the most specific of the three. It named Square’s processing rate (“typically $2.6% + $0.10 per transaction”), wrote that “you will still be responsible for standard merchant processing fees (usually around 2.5% to 3.5% + $0.30),” and named ChowNow’s pricing as “a flat monthly subscription fee starting around $119/month instead of per-order commissions.” Those numbers are claims I would still verify on the vendor’s own pricing page before acting — the engines run from crawls that can be one release cycle behind — but the caveats themselves are right.

The brief for this dispatch was that I would catch the engines repeating vendor marketing copy. The brief turned out to be wrong, and the better story is the inverse: on the commission-free question, the engines were more honest than I gave them credit for. If your operator instinct says “ChatGPT will sell me on something it doesn’t understand,” that instinct was off-target for this query in this week. The mechanism behind that surprise is worth naming: AI assistants train on the same restaurant-trade coverage that has spent the last three years writing “here is what commission-free actually means” explainers. The caveat has soaked into the source pool, so the engines now reproduce it.

The actual thin spots in this run are different, and they are quieter. They are worth naming because they are the kinds of things an operator might miss inside an otherwise-honest answer.

Google AI named Canva as a website-builder option for restaurants. Canva is a graphic-design tool with a basic website builder bolted on; it is not a serious option for a restaurant that needs a menu page, an online-ordering integration, or restaurant-specific schema. The fact that it surfaced here is a tell that Google AI lifted from a generic small-business roundup rather than a restaurant-specific one. If you are reading the Google AI answer and Canva looks like a contender, treat it as a flag, not a recommendation.

Perplexity recommended Shopify for restaurants — with a real-but-narrow framing. “Best only if your restaurant is really acting like a retail business too, such as selling packaged goods or merch online,” it wrote. The framing is technically correct, but it is also the kind of conditional an operator can miss while skimming. Shopify is not a restaurant website platform; it is a retail e-commerce platform, and using it for a restaurant means rebuilding the menu, hours, and ordering flow on top of e-commerce infrastructure that wasn’t built for any of those.

Every specific dollar figure the engines named is a same-day-verify item. The $119/month ChowNow figure, the $2.6% + $0.10 Square rate, the 2–4% processing-fee band — each one of those is a vendor-published number that can move on the vendor’s schedule, not the engine’s. If your decision turns on the specific price, open the vendor’s own pricing page and read the number today. The engines are useful as a shortlist generator; they are not a current-pricing oracle.

How an independent restaurant (or builder) gets named in these answers

If you read the methodology and the citation pattern above carefully, the path to being named in the next answer is already visible. The engines reach for sources they have learned to trust, lift the most-extractable sentence those sources offer, and assemble an answer that names three to five candidates. To be one of those candidates — whether the engine is recommending a website platform or a restaurant — you need to be readable, corroborated, and current.

The decision tree below walks the four checks an operator can run on a single sitting. Stop at the first “no” — that is the move to make this week. The teal verdicts are passes; the rust ones are the moves still owed; the warning at the bottom is the most common mistake operators make, a category of moves that feel productive and do almost nothing.

The first four moves compound. A restaurant that has done all four will be named in answers about platforms, neighborhoods, cuisines, and dietary fits — not because the model has been told to name it, but because the model has consistent, legible, corroborated text to lift. A builder who has done the same four for a client’s site is doing the same job at a different scale. The mechanism is the same; only the entity changes.

If you are reading this as the operator and your reaction is “I already do most of these,” that is the right read. Most operators I work with are further along than they think. The gap that costs an answer is usually one of these four, not all four. Pull the one you have not closed and close that one first.

The honest take

This dispatch is a snapshot, not a study. The engines change their weighting often. The prompts I ran are operator-shaped, not exhaustive. A different week, a different framing, and the names that came back would have shifted — that is part of the point of stamping a date on this and republishing the experiment when the picture changes materially.

The surprise of this run was the commission-free check. I expected the engines to repeat vendor marketing copy with the fine print stripped, and they did not. All three named the processing-fee caveat in their own words. That is genuinely good news for an operator reading those answers, and it is worth saying clearly: in late May 2026, on this question, the AI engines were doing better at “naming the math underneath” than I had braced for. That can change. It may already have changed by the time you read this. But it is what happened on the morning of May 30, and pretending otherwise would be the wrong kind of editorial framing.

The pattern underneath the snapshot, though, is the one I trust regardless. AI search engines are recommending restaurant website platforms the same way they are recommending restaurants: by reading a small handful of sources they have learned to trust, lifting the most-extractable sentence those sources offer, and assembling an answer that names three to five candidates. The platforms that get named are the ones whose framing has been most thoroughly repeated by the sources the engines already trust. The restaurants that get named work the same way. If you want to be one of either, the moves are the four above. There is no shortcut.

The reflex move — pick whichever platform the AI named first — is still the wrong move, even when the answer is honest. It is wrong because the engines are not weighing your operator economics. They are summarizing prose. The reason to read these answers is to see what the diner sees about you, not to take the recommendation at face value. The recommendation is a starting point. The four moves above are the work.

Frequently asked questions

What website platform does ChatGPT recommend for restaurants? On May 30, 2026, ChatGPT (signed-out, default mode) recommended Squarespace as the “best overall” for most independent restaurants, with Wix, WordPress, Toast Websites, and BentoBox rounding out the shortlist. For commission-free direct ordering specifically, it named Toast, Square for Restaurants, ChowNow, Owner.com, SpotOn Restaurant, and GloriaFood. The exact names move from session to session, and the recommendation is a starting point, not a verdict — the operator decision still comes down to ownership, cost, and how cleanly the model can read the resulting site.

Do AI search engines agree on the best restaurant website platform? Partially. On the May 30, 2026 run of the experiment, five platforms were named by all three engines tested — Squarespace, Wix, Toast, Square, and ChowNow. Four more were named by two of three. The remaining four singletons each told a different story about which sources the engine had absorbed. The divergence is the more useful signal: it tells you which engine is reading which sources, which is the operator-relevant question if you want to be named.

How do I get my restaurant cited in AI answers? Own a plain-text answer to the question a diner would ask, mark it up with restaurant schema, keep reviews fresh, and earn corroboration from sources the engines already trust. The full mechanism is in the library piece on how to get cited in Google’s AI Overviews; the same moves carry to ChatGPT and Perplexity because all three engines look for the same kind of legibility. The four-number visibility check is in the companion dispatch on appearing in AI search.