What makes ChatGPT treat our content as expert rather than marketing?

Your website reads well. The agency that wrote it was good. The case studies are polished, the service pages are persuasive, and the blog has a hundred posts with sensible titles. Yet when a buyer asks ChatGPT, Gemini or Claude who understands their problem, the assistant names a smaller competitor whose website looks like it was built in a weekend.

Good writing and expert content are different things. A page can be beautifully written and still be marketing copy: text designed to move a reader towards a decision rather than help them understand something. In a Princeton and IIT Delhi study presented at KDD 2024, adding quotations from relevant people raised a source's visibility in AI-generated answers by about 41%, adding statistics by about 34% and citing sources by about 30%. Rewriting the same text in a more authoritative tone raised it by only about 10%. It was tested on a research engine, so treat it as directional, but substance moved the result far more than style.

The competitor with the weekend website has, somewhere on it, a partner writing frankly about what goes wrong, what it costs and when their approach does not apply. This guide sets out the signals, a phrasing table your writers can use, how to apply it to service pages, and a test you can run this afternoon.

#What is the difference between marketing copy and expert content?

Marketing copy asks "how do I get this person to believe X". Expert content asks "what does this person need to know to judge X for themselves". Everything else follows from that.

Most business content is marketing, even when it does not feel like it. It is written from the seller's side of the desk: benefits, nothing awkward, a call to action. "Our proven approach delivers measurable results" tells a reader, or an assistant, nothing except that a firm would like to be hired.

Expert content is written from the reader's side: how something works, where it goes wrong and what it depends on. "For firms under about £5m, a full audit is usually unnecessary and a review will satisfy most lenders; above that the picture changes, and here is why" is expert content. It helps, and only someone who knows could have written it.

#What signals tell an assistant a page was written by an expert?

Five things, in rough order of importance: something original, something specific, an honest account of limits, visible reasoning, and a tone that treats the reader as a peer.

#It says something new

An expert notices what a summariser would not: that clients always ask the same question in month three, or that the standard advice suits large firms and fails small ones. Google's guide to its generative AI features (updated July 2026) makes the same distinction, contrasting "commodity content" that "could originate from anyone" with first-hand experience.

#It is specific

Expert content has numbers, cases and named sources: "a food manufacturer with three sites and about 200 staff", not "a client of ours". Where a figure has a source, name it. Where it is an observation, say so. Where you have none, do not invent one.

#It admits limits

Compare "our approach delivers a 40 per cent improvement for every client" with "the improvement varies a great deal and depends mostly on how organised the firm was to start with". The second is more useful and more credible, because it reads like someone who has watched it happen.

#It shows its reasoning

The page defines the problem, says why the answer works, names the alternatives and why they are less good, and says how it can fail.

#It treats the reader as a peer

No urgency and no fear: "this works for X but not for Y" instead of "this is the answer". Our guide to whether ChatGPT and Google quote AI-written or keyword-led content shows the opposite patterns side by side.

#What do AI assistants actually pick up on?

Substance: quotations, figures and cited sources, far more than tone. The Princeton and IIT Delhi test measured how much each change to a page moved its visibility in AI-generated answers:

A 2025 study of ChatGPT citations points the same way, though it shows correlation, not cause. SE Ranking found in November 2025 that pages with expert quotes averaged 4.1 citations against 2.4 without, and pages with 19 or more data points averaged 5.4 against 2.8 for pages with little data.

Search-connected assistants such as ChatGPT search, Perplexity and Google's AI features retrieve a handful of pages, compare them and lift the passages that answer best. They do not verify each figure. Buyers do: in the TrustRadius 2026 B2B Buying Disconnect Report (July 2026), 94% of technology buyers who used AI said they fact-checked its answers at least sometimes. An unsupported figure may get past an assistant. It does not get past a buyer who clicks through.

#Does the author's name matter to ChatGPT?

It matters to Google and to readers; for ChatGPT, no published study shows it checks bylines directly. The benefit is indirect, and it comes through two routes.

The first is Google. Its guidance on creating helpful content asks "Do pages carry a byline, where one might be expected?" and notes that "E-E-A-T itself isn't a specific ranking factor". Because Google's AI features, and assistants that draw on Bing, start from a search index, what helps a page rank helps it get retrieved, though ranking does not guarantee a mention.

The second is corroboration: a partner whose name and subject appear consistently on your site, on LinkedIn and in the trade press is easier to connect to that subject. LinkedIn is itself a source: a Semrush study (March 2026) found that in ChatGPT search and Google AI Mode, 59% of LinkedIn citations came from individual members rather than company pages. Why AI assistants care about E-E-A-T explains the idea in plain terms.

#Why is our well-written content not getting us named?

Usually because it was written to persuade, by someone who did not do the work, and nobody with expertise put their name to it.

For example, a property consultancy's page on rent reviews explains what a rent review is and that the firm can help. It has no author, no example, and nothing on what goes wrong or when a tenant should not bother contesting. To an assistant it is one of a hundred similar pages.

ID-Pal, an identity verification firm that had bought NorthRow and folded it into id-pal.com, had a version of this problem. Its case study puts it plainly: "AI could describe ID-Pal. It never recommended it." It records: "Across 36 recorded AI answers to the questions ID-Pal's buyers actually ask, ID-Pal was named three times, and only for its brand or its Salesforce integration. AI knew who it was, and had no reason to recommend it."

Two things changed. One was technical: "About four hours of structured data work declaring NorthRow part of ID-Pal on both domains." The other was content: "Five industry pages rebuilt: fund administration, wealth management, legal, accountancy and estate agency. Each opens with the question a compliance lead types, names the regulator, and answers inside the first ninety words." The result the ID-Pal case study reports: "UK buyer questions where AI names ID-Pal: none in May, two in August, four in September 2026."

The fix is rarely to write more. It is to put the expertise the firm already has onto the page. If outside writers produce your content, how to outsource content without losing your voice covers how to get the expertise in.

#Which phrases read as marketing, and what do experts write instead?

The pairs are illustrative. For each phrase on the left, ask what you would say to a client across the table, and write that.

The G2 figure comes from G2's "Answer Economy" research of 1,076 software buyers. Every figure in expert phrasing needs a source like that.

#How do we make service pages less salesy without hurting conversions?

Answer the buyer's question in the first paragraph, show how the service works and where it does not fit, then put the proof and the call to action after the substance.

Service pages matter here because, in our experience, "who should we hire" questions are answered from service pages, comparison pages, directories and reviews more often than from blog posts. Example of a rewritten opening:

• Before: "We are a leading provider of outsourced payroll services. Our experienced team delivers accurate, compliant payroll for businesses of all sizes, so you can focus on what matters. Contact us today."
• After: "We run payroll for UK firms with 20 to 250 staff that have outgrown doing it in-house. Most come to us after a missed pension deadline or a key person leaving. For the first month we run your payroll alongside your old system, so any gap in the data shows up before it costs anyone money. If you have fewer than about ten staff, payroll software is usually enough, and we will say so."

The second names who it is for, why they buy, how it works and when not to buy, and it still ends in a call to action.

Comparison pages follow the same rules. For Helm Club, a founder peer group: "We built the structured data AI engines read, then a comparison page for every question a founder asks when weighing up peer groups: Helm against YPO, Vistage and the rest." And: "Every comparison page we have published ranks first for its question on Google and the AI engines. The pages not yet built are the ones where a competitor still writes the answer." The Helm Club case study has the detail.

#How do we write content that reads as expert?

Start from what you know that others do not, be specific, say where it does not apply, explain why, and leave out the sell.

#Step 1: Find the observation

Ask the person who knows the subject what clients get wrong and what the standard advice misses.

#Step 2: Add the specifics

The number, the case, the timeframe and the source. Anonymise clients but keep sector, size and outcome.

#Step 3: Write the limits

When this does not apply, what it depends on, what you are not sure of.

#Step 4: Show the reasoning

For each claim, say why, and say why you prefer this option to the alternatives.

#Step 5: Take out the sell

Delete superlatives, urgency and the closing pitch; end with one calm sentence about where to go next.

#Step 6: Put a name on it

Author, role and why they know this subject, consistent with their LinkedIn profile.

#How do we test whether a page is expert content?

Read it and ask five questions. The more often the answer is the first option, the more expert the page.

1. Does it contain a thought someone else could not have had, or could anyone in the industry have written it?
2. Are the claims specific, with figures, cases and sources, or general assertions?
3. Does it say when the advice does not apply?
4. Can a reader follow the reasoning step by step?
5. Does it inform a peer, or persuade a prospect?

In our experience most firms find their pages fail the first and third questions almost everywhere. The full method for scoring a whole site is in how to audit existing content for AI assistants, and why ChatGPT never quotes most blogs covers the commonest failure.

#What to do this month

1. Run the five-question test on the five pages that matter most, starting with service pages.
2. Spend fifteen minutes per page with the partner who knows the subject: one observation, one example, one limit.
3. Rewrite each opening so the answer and the observation come first and the sell is gone.
4. Give your writers the phrasing table to apply line by line.
5. Add a named author with role and biography to every page that lacks one, matching their LinkedIn profile.

#Frequently asked questions

#Does ChatGPT check who wrote an article before quoting it?

There is no published evidence that it does. Authorship helps indirectly, through Google and through a name that appears consistently on your site, on LinkedIn and in the trade press. Put a real name on every page anyway, because buyers look.

#Does content written with AI count as expert content for Google and ChatGPT?

It can, if the expertise is on the page. Google's guidance on AI-generated content says "Using AI doesn't give content any special gains. It's just content." An AI draft becomes expert content when the person who did the work adds the observation, the specifics and the limits.

#Should our articles cite other firms' research or only our own?

Both. In the Princeton and IIT Delhi test, citing credible sources raised visibility by about 30%, and your own data gives an assistant something it cannot find elsewhere. Always name the original source and its date. Whether publishing your own research helps covers producing it on a small budget.

#How many case studies does our website need to look credible to AI assistants?

No study gives a number. Our view: one for each service and sector you want to be named for, each leading with a result. arum's rebuilt site is an example: "Each of the 12 case studies leads with a number, and each of the nine sector pages names the regulation in play." See the arum case study.

#Does admitting limitations make us look weak to buyers?

The opposite, in our experience. A buyer who reads "this approach does not suit firms with fewer than ten staff" trusts the rest of the page more, because the firm has shown it will tell the truth. Length is a separate question: whether ChatGPT prefers long or short articles covers it.

The expertise is already in your firm; the job is getting it onto the page. If you want an outside view of how your pages read, book a free Growth Review: 30 minutes with David Adams and three fixes in writing. If the site itself needs rebuilding around buyers' questions, that is Website That Sells.

#Sources