Making a page easy for AI systems to read is now the entry requirement, not the advantage. Clear headings, a direct answer near the top, self-contained sections, tables, schema and clean server-rendered HTML all help an assistant find and lift the right passage, and any competent competitor can copy that structure in an afternoon. What decides whether an assistant quotes your page or silently folds it into a general answer is whether the page contains something the other sources do not: your data, your results, your method, your interpretation. This article is about how to build that, where to find it inside a company that thinks it has nothing original to say, and how to reinforce it until the idea is associated with your brand.
What changed between ranking a page and being cited in an answer
A traditional results page lists documents and lets the reader choose. An AI assistant retrieves passages from several documents and writes one answer. That difference in output changes which pages get credit.
When an assistant gathers ten passages that all state the same fact, that fact becomes background knowledge. The answer can state it without crediting anyone, or credit whichever source is most convenient. When one passage contains a figure, a result or a conclusion the others lack, the sentence built on it has exactly one possible source, and a citation is the natural way to present it.
Two pieces of public evidence support planning around this. Google’s guidance on creating helpful, reliable, people-first content asks site owners directly whether their content offers original information, reporting, research or analysis, and whether it adds substantial value compared with other pages in the results. Separately, the 2023 GEO research paper by Aggarwal and colleagues, later published at KDD 2024, tested changes to web pages against AI-generated answers and found that adding statistics, direct quotations and cited sources raised a page’s visibility in those answers by up to about 40% in their tests. Keyword stuffing did not help and sometimes hurt.
Neither of those is a documented preference for opinion. Both point the same way: specific, attributable, verifiable material gets used. The mechanics behind it are covered in retrieval-augmented generation and query fan-out, and the content side in information gain.
Structure is the entry requirement
None of this is an argument against structure. An assistant cannot cite a passage it cannot locate, so the structural work still has to be done. A page that is ready for AI search generally has:
- A title that matches the question people actually ask
- A direct answer in the first paragraph, before any background
- Descriptive H2 and H3 headings rather than teasers
- Sections that make sense when read alone, because that is often how they are retrieved
- Figures placed next to the claims they support, not in a separate summary
- Named companies, products, versions and markets instead of vague categories
- Definitions for the terms the page depends on
- Question-led sections that reflect real queries
- Sources a reader can check
- Content present in the server HTML, since most AI crawlers do not run JavaScript
Do all of that and an assistant can understand and extract your page. It still has no particular reason to prefer it. Structure makes content usable. Original material makes it worth using.
SEO, AEO, GEO and the layer above them
It helps to separate the four jobs, because they fail for different reasons and are measured differently. The first three are standard industry terms. Treating point of view as a distinct fourth layer is Brilad’s framing, not an established category.
| Layer | Job | Fails when |
|---|---|---|
| SEO | Get the page crawled, indexed and ranked | Crawlers are blocked, content renders only in the browser, or no page targets the query |
| AEO | Become the answer shown on the results page | The page ranks but no passage answers the question in the format the query shows |
| GEO | Be named and cited inside AI-written answers | AI crawlers cannot reach the page, or the brand is described inconsistently across the web |
| Point of view | Give the answer a reason to name you rather than anyone else | The page is well built and well ranked but says what ten other pages already say |
The first three can all be done correctly and still produce a page that reads like every other page on the topic. The differences between the first three are set out in GEO vs SEO.
Your unique right to answer
Before writing anything, answer one question about the topic: why should this company be the source that answers it? If the honest answer is that you read the same three industry reports as everyone else, the page will reflect that.
A right to answer usually comes from something operational rather than editorial.
| Source of the right to answer | What it can produce |
|---|---|
| Transactional data | Real distributions: what customers actually buy together, when demand moves, what gets returned and why |
| Delivery experience | What breaks during implementation, how long steps really take, which recommended approach fails in practice |
| Customer conversations | The questions buyers ask before signing, and the ones they ask afterwards |
| Support history | The misunderstandings that recur, which say more about a market than any survey |
| Technical depth | Measured results from your own tests on your own systems |
| Narrow specialisation | Detail a generalist cannot reach, for one segment or one geography |
| Length of time in the market | What changed, what reversed, and which predictions did not hold |
The practical difference this makes is easy to see side by side. Suppose two companies publish on the same category. The first writes that buyers are increasingly price-sensitive, cites a well-known industry forecast, and lists the same five vendors as everyone else. The second writes that across its last 150 sales conversations, three requirements came up more often than price, names them, and explains which kinds of buyer raised each one.
The second is more useful to a reader, and it is the only one of the two that creates something an assistant can attribute. The figure is specific, the source is identifiable and the claim is checkable in principle. That example is illustrative: the point is the shape of the claim, not the number. Publishing a number you did not measure is worse than publishing nothing, because a reader or a journalist who checks it will find it is unsupported. More on this in first-party evidence and original research.
Where to find a point of view inside your own company
Keyword tools report what has already been asked often enough to measure. They cannot tell you what your team knows. That material is usually sitting with the people who do the work, and the way to get it is to ask them directly.
These questions produce more usable content than a keyword list:
- What do customers most often misunderstand about this?
- What are we seeing now that the market will notice in a year?
- What changed in the last twelve months that we had to adjust for?
- Which assumption that everyone repeats is no longer true?
- What surprised us on a recent project?
- Which mistake do companies in this market keep making?
- What do our competitors seem to be ignoring?
- What advice did we give two years ago that we would not give today, and why?
Record the answers rather than asking for written drafts. Specialists who will not write an article will happily talk for thirty minutes, and a transcript contains the specifics that matter. The editorial work is then turning each answer into a claim with evidence attached.
That last step is where most attempts stop too early. “AI adoption is increasing” is a background fact available everywhere. “Adoption is increasing, and in our implementations the constraint has shifted from model cost to data access, which is why projects now stall at the integration stage” is a position. It can be argued with, which is what makes it worth quoting. Naming the person behind it matters too, as covered in E-E-A-T for AI search.
One article does not establish a position
A single page with a strong claim does not make a company the authority on it. Assistants assemble their picture of a brand from many sources, and a claim that appears in one place, on your own website, carries less weight than the same underlying idea appearing in several independent forms.
The useful approach is to treat an idea as a programme rather than a post. If a company identifies a real pattern in its market, that one idea can support a year of publishing, as long as each piece adds something instead of restating the article.
| Surface | What it should add |
|---|---|
| The main article | The full case: method, data, limits, conclusion |
| An expert’s own post | The judgement behind the finding, in their voice, under their name |
| A customer case study | One instance of the pattern, with the client’s own numbers and permission |
| A talk or webinar | Questions from the audience, and the answers, which often expose what the article skipped |
| A podcast appearance | Association of the idea with a named person on someone else’s platform |
| Industry press coverage | A third party stating the finding, which is the strongest form |
| Directories and review sites | Consistent description of what the company does, so retrieval connects these sources to one entity |
What does not work is publishing the same text in seven places. Duplicated content adds no new evidence, and on your own site it creates pages competing for the same query. Each piece needs its own angle, its own examples or its own audience.
This is the part that sits outside the website, which is why it belongs to GEO work rather than page optimisation. The supporting detail is in brand mentions and co-occurrence, digital PR for AI visibility and getting listed on the review sites AI assistants cite. A topic covered properly across many connected pages also builds the structure described in topical authority in the AI era.
Make the insight extractable
Original material still has to survive retrieval. Assistants work with passages, not whole documents, and a model reading a long page attends unevenly across it. An insight buried in paragraph fourteen, split across three sections, with its supporting figure in a summary at the end, may never be retrieved as one usable unit.
Practical rules that follow from this:
- Put the claim and the number that supports it in the same short section
- Give that section a heading that states the finding, not a label like “Our research”
- Write the section so it stands alone, without pronouns pointing at earlier paragraphs
- Say who produced the finding and over what period, inside the section
- Cut the warm-up sentences before the point
The retrieval behaviour behind these rules is covered in context windows and token budgets, content chunking for AI retrieval and quotability, which is Brilad’s name for writing passages that make sense when quoted alone. For worked rewrites, see how to write content that ChatGPT quotes.
What to measure
Rankings and sessions do not capture whether an assistant names you, and they cannot tell you whether what it says about you is correct. A measurement set for this work looks different.
| Signal | How to measure it | What it tells you |
|---|---|---|
| Mention rate | A fixed set of buyer questions, run repeatedly on each assistant | Whether you are in the answer at all, and how reliably |
| Cited URLs | Record which of your pages are linked, not just that the brand appears | Which content is doing the work |
| Competitive set | Which companies are named alongside you | Which market the assistant thinks you are in |
| Description accuracy | Read what the assistant says you do | Usually traceable to one outdated page, and fixable |
| Topic coverage | Which subjects produce citations and which never do | Where the next article should go |
| Share of voice | Your mention rate against competitors on the same question set | Direction of travel over months |
| Branded demand | Branded impressions in Search Console, and enquiries that name a specific article | Whether visibility is reaching buyers |
Answers vary between runs, between models and between signed-in users, so one screenshot proves very little and a fixed prompt set run on a schedule proves something. The method is in tracking brand mentions in ChatGPT, Perplexity and Gemini, with reporting in an AI visibility dashboard and the competitive view in competitor analysis in AI answers. A Brilad AI Search Audit sets this baseline before any content work starts, so later changes can be compared against it.
What a point of view is not
The idea is easy to misread, and three misreadings waste real effort.
It is not disagreeing with the industry. Contrarianism for its own sake produces claims you cannot support. An overlooked pattern, a better explanation of a familiar trend, or a documented correction all add information without requiring you to be against anything.
It is not an opinion without evidence. Assistants do not verify claims, so an unsupported assertion may well be repeated. That is a liability rather than a win, because the first journalist, analyst or buyer who checks it finds nothing behind it. Every position needs its method and its limits stated.
It is not a substitute for the technical work. If AI crawlers cannot reach the page, or the content only appears after JavaScript runs, the best original research in your market is invisible. The seven reasons an assistant does not mention a company are mostly access and identity problems, not content problems.
One more constraint is worth stating plainly. A language model writing from its training data reproduces what is already common, which is the opposite of what this article argues for. AI tools are useful for drafting, editing and checking once you have the material, and they cannot supply the material itself.
The sequence we use
For an article intended to be cited rather than merely to rank, Brilad works in this order:
- Start from a question a real buyer asks, in their words
- Answer it directly in the first paragraph
- Add the finding, data point or judgement that only this company can supply
- Show the evidence, including the method and the sample
- Have a named expert say what it means and why it matters
- Give specific examples, with names, numbers and dates
- Cite the external sources a reader would want to check
- Reinforce the idea across other surfaces over the following weeks
Steps one, two, six and seven are the structural work most teams already do. Steps three, four and five are where differentiation comes from, and step eight is what turns one article into a position.
Frequently asked questions
We are a small company with no research budget. Can we still do this?
Yes, and the cheapest material is usually the best. Twenty client questions written down, the results of one project measured before and after, or a correction of a claim your market repeats all count as original information. Formal surveys are the most expensive form, not the only one.
Does publishing our data help our competitors?
Some of it would, which is why the judgement call is about which portions to publish. Aggregated findings, ranges and patterns can usually be shared without exposing anything commercially sensitive, and client data needs permission regardless. Competitors can copy a published finding, but they cannot become the source of it.
Will AI assistants actually credit us?
Nobody outside the AI companies controls their answers, and no agency can guarantee a citation. What you can do is make yourself the only available source for specific claims, make those claims easy to retrieve, and measure the result on a fixed question set over time.
How long before this shows up in AI answers?
Assistants that search the web live can reflect a new page once the relevant crawler revisits it, which is often days or weeks. What a model knows without searching only changes when it is retrained, which takes months and is outside anyone’s control. Corroboration across other sites accumulates on its own timetable, usually slower than publishing.
Can we just add statistics to our existing pages?
Adding figures you have actually measured, next to the claims they support, is worth doing on pages that already rank. Adding figures copied from other sites is not, because those numbers already exist everywhere and credit flows to whoever published them first.
Does this conflict with writing for search engines?
No. The structural checklist earlier in this article is standard SEO and AEO practice, and it still applies. The argument is about what goes inside that structure, not about replacing it.
Sources
- Google Search Central, “Creating helpful, reliable, people-first content”, including its self-assessment questions on original information, reporting, research and analysis
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, “GEO: Generative Engine Optimization”, 2023, published at KDD 2024
- Google Search Central documentation on AI Mode and query fan-out
- United States patent granted to Google in 2022 describing the scoring of documents by information gain, treated here as evidence that the idea has been studied rather than as a confirmed ranking signal