ChatGPT's premium model cites brand sites 7x more. Your homepage is now the front line.

When ChatGPT 5.4 Thinking quietly became the premium default in mid-March, something broke in the AI-search playbook most agencies had been selling. Forbes mentions, TechRadar round-ups, Reddit threads -- the whole "earn third-party coverage and you'll be cited" strategy -- stopped paying off the way it used to. The model is now reaching for brand websites first.

A study published by Writesonic in March, covering 1,161 citations across 50 prompts and 119 conversations, put a number on the gap. GPT-5.4 Thinking sent 56% of its citations to brand-owned domains. GPT-5.3 Instant, the free default, sent 8%. Search Engine Journal's coverage confirmed the same numbers and noted the two models share only 7% of cited sources -- effectively two different search engines wearing the same logo.

How the new model actually searches

The mechanic matters more than the headline. GPT-5.4 averaged 8.5 sub-queries per question and used the site: operator in 156 of its 423 total queries. That means it isn't waiting for your business to surface in a generic web search. It's specifically asking Google "what does site:yourbusiness.ie say about delivery times?" and reading whatever your own pages return.

The breakdown of where those brand-site citations land is just as telling: 22% homepage, 19% pricing page, 10% product page. And 75% of GPT-5.4's citations don't appear in standard Google or Bing results at all -- so traditional rankings are no longer a reliable proxy for AI visibility.

For premium users, your homepage and pricing page are the interview. The model is asking them direct questions and quoting the answers.

Why this hits Irish SMEs harder

Most Irish SME websites we audit fail the new test for one of three reasons.

• Pricing lives in a PDF, an email, or a "request a quote" form. If a model runs site:yourrestaurant.ie lunch menu price and gets nothing parseable, it cites a competitor that publishes its menu in plain HTML.
• The homepage is a hero image and a tagline. "Crafted with care since 1998" tells the model nothing. It needs the city you serve, the products you sell, the hours you open.
• Product and service pages are rendered in JavaScript. A separate analysis on PPC.land noted that JS-dependent pricing tables and lazy-loaded content are simply invisible to the crawlers behind these models. A React-rendered Squarespace pricing block disappears.

This is on top of an already-narrowing citation surface. Reporting on a Resoneo study of 27,000 ChatGPT responses found that the average number of unique domains per answer dropped from 19 to 15 after the GPT-5.3 transition -- a 20% cut. Fewer slots, more competition, and a model that now prefers your own site over the press coverage you worked to earn.

What to do this week

The fix isn't a new content strategy. It's making the pages you already have legible to a machine that's reading them directly.

• Put pricing on a real page. Server-rendered HTML, with numbers and units the crawler can read. "From €45" beats "Get in touch for a quote" every time the model fans out a comparison query.
• Rewrite the homepage as if a stranger had to summarise your business in one sentence. What you do, who you do it for, where you do it. Schema.org LocalBusiness or Organization markup gives the model the same facts in a structured form it can quote verbatim.
• Audit your product and service pages for JavaScript dependence. Disable JS in your browser and load the page. If the prices, descriptions, or hours vanish, GPT-5.4's crawler sees the same blank space.
• Add JSON-LD for the things you want quoted. Products, FAQs, opening hours, service areas. A small Galway hotel that publishes Hotel + Offer markup is dramatically more quotable than one that buries the same facts in a marketing paragraph.

None of this is exotic. It's the same hygiene that's been good practice for a decade -- but the cost of skipping it just changed. Pre-2026, a quiet homepage cost you a few rich-result snippets. In April 2026, it costs you the answer slot inside a ChatGPT conversation that your customer will never tell you they had.

If you want to see what's currently parseable on your own domain, the Baseline schema generator is the fastest way to find out. It's free, and it tells you exactly which pages need the most work.