Competitor research with AI: reading rival hotels' websites, reviews and offers

AI competitor analysis means using an AI assistant to read and compare the public text rival hotels publish: their room pages, offers and recent reviews, with that text pasted in as the source. The assistant summarizes and tabulates; a person checks every row against the link and the date.

A sales manager asks an assistant what the hotel down the road offers and what it charges. The answer comes back tidy and confident, and part of it is invented: a spa that closed, a rate from two seasons ago, a package nobody runs. This article shows how to do the same research so that nothing invented reaches your decisions: what to collect from each rival, how to prompt for a table, how to test every finding, and how the findings become copy, offers and service fixes.

Why does an AI assistant invent a rival's facilities and prices?#

Because it is writing likely text, not reading the rival's website. Unless the source text is in front of it, every detail about a rival hotel is a guess that reads like a fact.

An assistant produces the most probable next words. A plausible room type and a plausible price look exactly like a correct one on the screen, and nothing in the sentence warns the reader which is which. Even when an assistant can open a link you give it, hotel pages often show rooms and rates only after a date search, so it sees the marketing shell, not the price for your dates.

Prices and packages move. A finding without a collection date is unusable within weeks, and nobody can tell later whether it was ever right. The failure is quiet: no error message, no blank field, just a confident sentence. That is why the method has to be mechanical rather than trusting.

The fix is to paste the text you want compared, keep the link and the date beside it, and ask the assistant to quote the sentence behind every claim it makes. Tell it explicitly to answer that the text does not say when the source is silent. Without that instruction, silence gets filled.

Reading versus recalling#

Recalling is what the assistant does from memory: producing text that sounds like what a hotel website would say. Reading is what it does with pasted source text: extracting what the page actually states. The two look identical in the output window, so the only reliable difference is the input. If the page was not pasted in, the assistant is recalling, and every facility, price and package name in the answer should be treated as fiction until checked.

What should you collect from each rival, and how should you prompt for it?#

Collect three kinds of public text per rival: the room and facility pages, the current offers and packages, and the recent guest reviews. Then hand that text to the assistant with a fixed output shape, so every rival is read the same way.

Room and facility pages give you what each room type promises in the hotel's own words: size, view, occupancy and what is included in the rate. Offers and packages give you the name, what is inside, the conditions, the blackout dates, and where the offer sits on the page; a package in the hero banner is a priority the hotel has chosen. Recent reviews, rather than the all-time average, tell you what guests experienced lately; record the platform and the date range, keep the comment text and drop the guest's name.

Also note the segment the page is written for: families, couples, business travelers, groups. The words a hotel chooses tell you who it is chasing, which is often more useful than its rate.

A prompt that returns a table, not an essay#

Write the prompt in six parts: the role, the pasted source, the exact task, the columns you want back, the quoting rule, and the instruction to say unknown when the text is silent. For example: "You are summarizing hotel web pages for a competitive review. I am pasting the room page of Rival A below. Extract each room type with its stated size, view, occupancy and inclusions. For every cell, quote the sentence it comes from. If the text does not say, write 'not stated'. Return a table with these columns: room type, size, view, occupancy, included."

The quoting rule is what makes the output checkable. An assistant that must quote a sentence for every claim either produces the sentence or reveals that it cannot. The "not stated" instruction is what stops it from filling silence with a guess.

What to leave out#

Leave out the all-time review average; it mixes years of different management. Leave out third-party descriptions of the hotel, because they are often outdated copies of each other. Leave out your own memory of the property, which is exactly the kind of unverified input the method exists to exclude. Keep one document per rival per collection date, so next month's version can be compared with this one instead of replacing it.

For the review side of this work, the same collection discipline applies when you turn hundreds of your own or a rival's comments into a short list: see reading guest reviews with AI.

How do you fit the comparison on one page?#

One row per rival and five columns: what they promise, what guests praise, what guests complain about, the gap you can use, and the source. One page forces choices; a long spreadsheet gets read once and then forgotten.

The promise column uses their words, shortened but not paraphrased. Your paraphrase is already an interpretation, and interpretations drift. The praise and complaint columns carry themes with a count of how many reviews mentioned each, so a single angry review does not become a strategy.

The gap column is your judgement, written as a sentence a manager could act on: not "bigger lobby", but "the lobby is the only place with wifi after checkout". The source column holds the link and the collection date. A row without a source is deleted, not softened.

Put your own hotel in the first row and fill it in the same way. A comparison that starts from your brochure flatters you by construction. Then read the finished page as a guest would: which row would make someone choose them, and which row would make someone choose you.

The example rows are illustrative, not real. Note the counts beside each theme: they are what separates a pattern from an anecdote.

How do you test a finding before you act on it?#

Four tests: can you point to the sentence, is it current, do several guests say it, and would a guest notice the difference. A finding that fails any of them stays out of the decisions.

The quote test: open the source and find the exact sentence. If the assistant cannot produce it, the finding is dropped rather than reworded. The date test: a rate or an offer without a collection date is history; reopen the live page before the finding reaches a meeting.

The volume test: one review is an anecdote. A theme needs several independent reviews saying the same thing in different words before it becomes a complaint pattern. The notice test: a gap only matters if a guest choosing between you and them would feel it; most differences in facilities are invisible at the moment of choice.

Add a second-reader test: someone who did not run the prompt reads the row and the source and says whether the row is fair. The person who wrote the prompt is the worst judge of it. Record the dropped findings and the reason; that list stops the next person from repeating the same work and shows where the assistant is weakest.

How do findings become copy, offers and service fixes?#

Sort each gap into one of three buckets: something you can say better, something you can package, and something you must fix before you say anything. Each row gets an owner and a date, or it stays a note.

Say better covers a true strength you never wrote down. Rewrite your own room and offer pages in your own words; never lift their sentences or their photographs. Package covers a combination rivals do not offer, built from what you already have: a late checkout, an airport transfer, a breakfast time that suits one segment.

Fix covers a complaint theme that is also true of you. Marketing a promise you cannot keep converts once and then costs you the review. Assign an owner and a date to every row; an unassigned finding is a note, and notes do not survive the week.

Feed the wording work into your copy process and the complaint themes into your fix list, so the same evidence is not collected twice. The wording side has its own guide: AI for hotel marketing copy. Repeat the collection on a fixed rhythm so the page shows movement: a promise that disappeared, a complaint that grew, an offer that was withdrawn.

What are the limits of AI competitor research?#

Public pages and public reviews only: no accounts, no automated collection against a site's terms, and no conversation with rivals about prices. The method reads words; it does not read the market.

Reading a public page is one thing. Creating an account to see member rates, or running a tool that collects pages at scale against a site's terms, is another. Check the terms and ask before automating anything. Do not ask rivals for their rates and do not exchange rate information with them, directly or through a third party; competition law treats price exchange between competitors as a serious matter, so take advice for your own case.

A price you see is an asking price for a date and a channel, not what the rival earns. Reviews are public, but the people who wrote them are not your data source: strip names and anything identifying before you paste review text into a tool. An assistant's summary is not evidence; the page is, so keep the source with the finding or the finding does not exist.

Rate comparison and index measures are different jobs. Do them in their own process, not inside a text-comparison prompt; see hotel competitive set and MPI, ARI and RGI.

What to do next#

Pick one rival today, paste their room page and a set of recent reviews into your assistant, and build a single row with the promise, the praise, the complaint, the gap and the source. Then run the quote test on it. That one row is the whole method in miniature: collect the public text, compare it in a fixed shape, check it against the source, and only then decide whether to say something better, package something new, or fix something first.