How to Stand Out When Everyone Uses AI (We Measured 24 Firms)

How to stand out when every competitor uses AI starts with a measurement. On 1 October 2026 we removed the names from 24 Dubai business setup homepages and asked an embedding model to match each page to the rest of its own text. The model found 4 to 8 firms by their own words. The other firms matched a competitor's page closer than their own.

AI did not make these firms worse writers. The firms all write fluent homepages, and every one serves its text cleanly to AI crawlers. Output is solved. What nobody has solved is the next step: giving a buyer, an AI answer, or an AI agent one reason to choose you that is not also true of the firm next door.

What did we find when we measured 24 firms in one category?

We measured thisWe measured business setup in Dubai because few categories are more crowded, and every firm in it sells the same regulated service: a trade licence, a visa, a bank account. The sample is the 24 firms named in web results and "best of" lists for business setup searches on 1 October 2026. We measured each homepage the way most AI crawlers read it: the initial HTML, without running JavaScript.

Sameness and agent readiness across 24 Dubai business setup homepages

Can a buyer tell these firms apart once the name is gone?

We measured thisWe measured each firm six ways in the name-removed test: two embedding models (BAAI's bge-small and bge-base) and three different ways of splitting each page into two halves. Every run asked the same question. Given the first half of a homepage, which of the 24 second halves sits closest to it?

Firms in this category all sell the same service, so the topic alone cannot pick out a firm's other half. Only what is particular to the firm can. Firms with their own way of speaking find their own other half. Firms writing the category average find whichever competitor wrote the average best.

The best run found 8 firms. The worst found 4. Guessing at random would find 1. Firms are not pure noise, but for at least two thirds of them the machine reading the page cannot hear the firm behind it.

We measured thisWe counted the phrases next. Of the 24 firms, 23 use "expert" on their homepage. In our data, "quick" or "fast" appears on 21, "trusted" on 20, "free consultation" on 20, and "100% ownership" on 17.

Of the 24 firms, 22 put some version of "business setup in Dubai" or "start your business" in their page heading. None of these phrases is false. Any competitor could say each one, which is exactly why none of them can be the reason a buyer chooses.

Anil Doshi and Oliver Hauser measured the same pull in a controlled experiment, published in Science Advances in 2024¹:

"Generative AI–enabled stories are more similar to each other than stories by humans alone." Anil Doshi and Oliver Hauser, Science Advances, 2024

Each writer improved. The group converged. Our 24 firms are that experiment run on a real market, with real money behind every page. The mechanism has its own page, why every brand sounds the same, and the category-level cost is in the sea of sameness.

Why is "add a human touch" the same advice too?

Articles on how to stand out all give one answer. We read four of the top results in full. Entrepreneur's September 2026 column says keep your personal stories human and break the AI formula. A Forbes Business Council post from February 2026 says use AI's efficiency to fund human strategy.

A Gulf trade piece in Communicate Online, written by a Heriot-Watt Dubai professor, tells CMOs to measure how semantically similar their assets are to their competitors'. An agency blog names taste as the new edge.

The advice on how to stand out is reasonable, and it has converged. Four sources, one answer, zero measurements. The Communicate Online piece comes closest to the real move by recommending that you measure similarity, and then publishes no measurement. Telling firms to "be more human" produces the same result as telling them to "be more expert": a category where every firm says it.

Sameness is not fixed by tone. A firm that adds a founder story to the same promises still makes the same promises. Firms stand apart only on a claim that would be false if a competitor made it.

Is being agent-ready a way to stand out?

In this category, no. These firms are already at parity on agent readiness. We found all 24 homepages serve their full text without JavaScript, the median at about 3,100 words. None of the 24 firms blocks GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, PerplexityBot, Google-Extended or CCBot in robots.txt.

And 18 of the 24 firms publish an llms.txt file, the summary page written for language models. At least 5 of those 18 files announce in their opening lines that an SEO plugin generated them: Rank Math on 3 sites, Yoast on 1, All in One SEO on 1.

Plugins are doing for agent readiness what language models did for copy. Firms get the basics by default, all at once, so the basics stop separating anyone. Cloudflare launched its Agent Readiness score² on 17 April 2026, and its own scan found only 4% of sites declare their AI usage preferences. By the time a scorecard is free, the score it measures is on its way to the floor.

We found the real gap one step further in, where an agent stops reading and starts acting. We measured 332 form fields across the 20 homepages that carry form fields in their initial HTML. Of those, 48 fields on 9 sites have no machine-readable name: no label, no accessible name, no placeholder.

Most are the dropdowns that matter for this service, such as nationality, business activity and licence type. An agent reading the page through its accessibility tree sees an unnamed menu where the buyer's most important answer goes. Our guide to how AI agents see your website walks through that tree and the 20-minute self-test.

Output is solved. Growth is not.

Production stopped being scarce for firms some time ago. Ahrefs analysed about 900,000 newly published pages³ and found that 74.2% contained AI-generated content. Firms can now ship a fluent page, a blog post, a set of FAQs, and a clean llms.txt file. When every firm can produce, production stops being the competition.

The choosing moved, too. G2 surveyed 1,076 B2B software buyers in March 2026⁴. In that survey, 51% now pick an AI chatbot over Google as their usual first stop for research, up from 29% a year earlier.

Per G2, 69% chose a different vendor than they planned because of what the chatbot said. Per the same survey, 33% bought from a vendor they had never heard of. G2 surveyed software buyers, not founders setting up companies in Dubai, but the direction holds for any service sold to a buyer who asks a machine first.

Agents sharpen this. Amine Allouah, Omar Besbes and colleagues ran shopping agents across product categories and reported in a 2025 arXiv paper⁵ that the agents showed "choice homogeneity": demand concentrated on a few "modal" products while others were ignored entirely.

An agent working down a shortlist does not reward the 24th fluent version of the same promise. It needs a reason, and it can only use the reasons you put on the page.

Growth is the part AI did not solve. Growth needs someone to decide what is worth saying, prove it, and check whether it moved a number. That work is what Growth Engineering names: one agreed growth target, and every page, post and test aimed at it.

Four tests for standing out when every competitor uses AI

1. The name-swap test. Put your closest competitor's name on your homepage heading and first paragraph. If every sentence is still true, you have given buyers and machines nothing to choose you for. In our sample, 22 of 24 headings are some version of "business setup in Dubai" or "start your business", a line every competitor can sign.

2. The checkable-number test. Put one claim in your page heading that a buyer can hold you to and a competitor cannot copy without lying. A licence in 3 days passes this test. "30,000+ entrepreneurs trust us", a heading in our data, is about the firm, not the buyer. Only 2 of 24 firms put a buyer outcome with a number in their heading.

3. The agent-completion test. Walk your enquiry form as software would. Every field needs a name an accessibility tree can read, above all the dropdowns where the buyer's real answer goes. 9 of 24 firms have at least one field an agent cannot name, most of them dropdowns.

4. The one-number test. Pick one growth number, such as qualified enquiries a month or cost per new client, and set a review date. Judge every page by whether it moved that number, not by whether it sounds human. A distinctive page that moves nothing is still decoration.

The first three tests make you choosable. The fourth makes the effort accountable. Without the fourth, the first three turn into another round of copywriting.

The limits of this study

Our study measures whether firms can be told apart, not whether the separable ones win more. We have not measured which of the 24 firms ChatGPT or Google recommends, so nothing here shows that the 4 to 8 firms the model could find get more business. That link is the next study, and until we measure it, we will not claim it.

The snapshot is one day, one category, homepages only. Firms whose services pages are sharper than their homepage score worse here than they deserve. The name-removed test measures how separable the words are, and a firm can still win on price, reviews or referrals with a generic homepage. The State of Sameness study on Dubai real estate found the same pattern from a different angle, through the production volume of one brokerage.

How this study was built

We took the firms named in web search results and "best of" lists for three business setup searches on 1 October 2026: 26 firms, of which we excluded 2 whose homepages lead with a different service. For each of the 24 firms, we fetched the homepage as plain HTML with no JavaScript, robots.txt and llms.txt.

We removed each firm's own name and brand words before the similarity test. The embedding models were BAAI/bge-small-en-v1.5 and bge-base-en-v1.5. Stock phrases were measured with fixed patterns across the visible text.

A field counted as unnamed only with no label, aria label, placeholder or title, and we excluded hidden spam-trap fields. Firms are not named because the pattern belongs to the category, not to any one firm. The raw data and scripts sit in our research record and can be re-run.