Is the most important information about your company stuck in the middle of the page, between a long introduction and a conclusion? AI is missing it. Models give far more weight to what sits at the beginning and the end of the text — everything in the middle ends up in a zone where the chance of being picked up is much lower. You're writing the right things in the wrong place. Moving your key information to the spots where AI actually looks can change everything — without touching the content.
You’ve written a perfect page. The brand is mentioned, the service is described, the CTA is clear. But the key information is in paragraph 5, at the heart of the page. For a human reader that’s not a problem — they scroll and find it. For an AI model it’s a serious problem, because the middle of the page is the dead zone.
This isn’t an opinion. It’s a measurable, documented behavior that in the research world is called the “lost in the middle effect”. And it has direct implications for who AI cites and who it ignores in its answers.
What Positional Encoding is and how it shapes what AI remembers
Transformer models don’t process text word by word like a human reading in sequence. They receive the entire text all at once, but they need to know where each word is. Positional encoding is the mechanism that provides this position information.
Minaee et al., in their 2025 survey, describe it very clearly:
“Positional encoding is incorporated to fuse information about the relative or absolute position of the tokens in the sequence.”
(Large Language Models: A Survey)
Translated: every token receives a kind of “address” that tells the model its position in the text. Without this mechanism, to the Transformer the words would be an unordered set — like Scrabble tiles dumped on the table. Positional encoding is what turns a pile of tokens into a sequence where “first paragraph” and “last paragraph” actually mean something.
And here comes the part that concerns you. Because not all positions carry the same weight.
The “lost in the middle” problem: it’s not theory, it’s measured
A group of researchers systematically tested what happens when you move the same piece of information to different spots in the text. The result, published in ACL Transactions, is unambiguous:
“We observe that performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models.”
(Liu et al., 2024)
That “even for explicitly long-context models” is the most striking part. Even models designed specifically to handle long texts — Claude with 200K tokens, GPT-4 with 128K — suffer from this effect. It’s not the limit of one specific model: it’s a consequence of the architecture itself.
The survey by Gao et al. (2024) on RAG confirms the problem from the retrieval side as well:
“Redundant information can interfere with the final generation of LLM, and overly long contexts can also lead LLM to the ‘Lost in the middle’ problem.”
(Retrieval-Augmented Generation for Large Language Models: A Survey)
Which means: even when an AI engine like Perplexity correctly retrieves your page, if the key information is in the middle of a long and redundant text, the model loses it anyway while generating the answer.
If the brand isn’t in the body text of the first and last paragraphs, as far as AI is concerned it doesn’t exist.
Where you put the brand changes everything: a test I ran
Before writing this article I took a real page — the “Services” page of a photovoltaic systems company — and ran an experiment with ChatGPT.
The original page was structured like this: 400 words of generic introduction about the photovoltaic market, then 600 words on the types of systems, then finally the company name and its specialization in paragraph 6, then 400 technical words, then the CTA.
I pasted the text and asked: “Summarize this page and tell me which company offers the service.” ChatGPT talked about photovoltaics in general. Of the company, not a single mention. The specific information was in the middle — lost.
Then I rewrote the same page, moving the brand and the differentiator. First sentence: the company name with what it does and for whom. Last sentence: the name again with the CTA. Same content, same 1,800 words, just repositioned.
Result: ChatGPT cited the company by name in the summary. Same content, different order, opposite result.
It’s not magic. It’s the “lost in the middle” effect: models make far better use of the information at the beginning and the end of the context than of the information in the middle.
Put the brand in the first 100 words of every page.
Why Italian pages are particularly vulnerable
I’ve noticed a pattern in the Italian companies I analyze. There’s a cultural tendency to “set the stage” before getting to the point — long introductions, historical context, preambles. In a magazine article that’s a virtue. For AI visibility it’s a problem, because you’re filling the first 200 tokens — the ones the model uses best — with generic content.
The “Wikipedia” introduction. Pages that start with “Photovoltaics is a technology that harnesses solar energy…” — 100 tokens wasted explaining something AI already knows better than you. Those tokens are the most valuable on the page and you’re using them for content the model has seen millions of times in the corpus.
The brand only in the header and footer. The company name appears in the logo, the menu and the copyright — elements that AI crawlers often classify as navigation, not content. If the brand isn’t in the body text of the first and last paragraphs, as far as AI is concerned it doesn’t exist.
The CTA with no context. “Contact us to learn more” as the last sentence. AI reads that final chunk carefully — and finds a generic sentence with no brand, no service, no differentiator. It’s the position with the second strongest signal on the page, and you’re wasting it.
What to do concretely
- Put the brand in the first 100 words of every page. Don’t wait until paragraph 5 to say who you are and what you do. The first mention of the brand together with the key service has to come right away — that’s where the positional signal is strongest.
- Repeat the key information in the last paragraph. The conclusion isn’t a summary for the reader — it’s a second chance for AI to capture the important information. Brand, service, differentiator: repeat them at the close in different words.
- Cut the generic introductions. If your “SEO Consulting” page opens with “SEO is a fundamental discipline in digital marketing…”, you’re sacrificing the most valuable position to say something that adds nothing for either the reader or AI.
- Hourglass structure: key information at the beginning, deep dive in the middle, key information at the end. The middle is for the humans who read everything. The beginning and the end are for the AI that extracts.
- Run the 150-token test: copy your page’s text, read only the first 150 tokens and the last 150. From those two blocks, is it clear who you are, what you do and why a customer should choose you? If the answer is no, AI has the same problem.
Positional encoding in the AI visibility chain
This mechanism combines with others. Tokenization decides whether your brand is recognized as an entity. Positional encoding decides whether that entity is “seen” based on where it sits in the text. The attention mechanism decides how much weight it’s given. And the context window limits how much text the model can process in total.
If your brand is well tokenized but buried in the middle of a long page, the “lost in the middle” effect penalizes it. It’s like having a perfect business card but handing it over when the other person is already distracted.
Reopen your 5 most important pages and move the brand, the main service and the CTA into the first 2 and the last 2 paragraphs. Lost in the middle isn’t a bug — it’s a consequence of the architecture. Those who know it, adapt.
