Why Pronoun-Heavy Sentences Are Harder for AI Systems to Extract Accurately
Pronoun-heavy sentences are harder for AI systems to extract accurately because AI search retrieves and quotes passages, not whole pages. A passage that says "it", "they" or "this" without naming the subject loses its meaning once it is lifted out of the page. Anthropic's research on retrieval gives a clear example. A chunk of text that says "The company's revenue grew by 3% over the previous quarter" does not say which company or which quarter. So it is hard to retrieve for a question about a specific company. Adding that missing context to each chunk reduced failed retrievals by 49% in Anthropic's tests. For writers, the fix is simple: name the subject in the sentences that carry key facts.
Why do pronouns cause problems?
- Passages travel alone. Search systems split pages into passages or chunks, find the ones that match a question, and pass them to the model. A passage that begins "It also supports..." has lost the name it refers to.
- Matching needs names. A question about "[Product] pricing" matches passages that contain "[Product]". A passage saying "Its pricing starts at $20" may not match at all.
- Wrong referents. When a page discusses several products or companies, "it" can be read as the wrong one. That produces answers that attach one product's features to another.
- Quotes need to stand alone. An answer quoting "They offer a 30-day free trial" tells the reader nothing about who "they" are.
Anthropic's approach to the problem, which it calls contextual retrieval, adds a short explanation to each chunk before indexing. Writers can do the equivalent at the source, by writing sentences that carry their own context.
Which sentences need the subject named?
Not every sentence. Focus on the ones most likely to be extracted:
- The first sentence of each section. Sections often become separate passages.
- Sentences stating a key fact: price, capability, limit, date, location, result.
- Sentences after a heading that is a question, since they form the answer.
- Captions, table cells and list items that could be read on their own.
- Any sentence on a page that discusses more than one product, company or person.
What does a rewrite look like?
Before:
"Our new plan launched in March. It includes unlimited projects and priority support. It costs $49 a month, and it integrates with Slack and Teams. This makes it ideal for growing teams."
After:
"The Growth plan for [Product] launched in March 2026. The Growth plan includes unlimited projects and priority support. [Product] Growth costs $49 a month and integrates with Slack and Microsoft Teams. It suits teams of 10 to 50 people who have outgrown the free plan."
The second version can be split anywhere and each part still says what it is about. It keeps one "it" where the subject is unmistakable, in the sentence after the subject was just named.
Are there other constructions that cause the same problem?
Yes, several:
- "This" and "that" without a noun: "This reduces costs" becomes "This scheduling feature reduces overtime costs".
- "As mentioned above": restate the point briefly instead.
- "The former" and "the latter": use the names.
- Headings without subjects, such as "Pricing" on a page covering three products. Use "[Product] Growth pricing".
- Tables without row labels, or with labels such as "Option A".
- Quotes without attribution in the same sentence.
How to keep content easy for AI to parse without making it robotic is covered in how content can be easy for AI to parse without hurting human readability.
Where should you stop?
Repeating the name in every sentence reads badly, and stuffing a brand or product name looks like manipulation. A reasonable balance:
- Name the subject at the start of each section and in each sentence that states a fact someone might quote.
- Use pronouns freely in the sentences between, where the subject is obvious from the one before.
- Read the page aloud. If it sounds like an advert, you have gone too far.
Google's guidance on its AI features says there is no need to break content into chunks for AI. Clear writing that holds up when excerpted is what matters, which is what this approach produces.
How do you check your pages?
- Pick your ten most important pages and read the first sentence of every section. Does each name its subject?
- Copy random paragraphs into a blank document. Can a reader tell what they are about?
- Ask assistants specific questions about facts on those pages, and see whether they attach the facts to the right product or company.
- Look for mix-ups in answers, such as one product's price given for another, which often trace to ambiguous passages.
What might an illustrative case look like?
Consider an illustrative software company with three products described on one comparison page. Its paragraphs used "it" and "this plan" heavily. AI answers regularly gave the mid-tier product's integrations for the entry product, and one assistant quoted "It includes single sign-on" for the wrong plan. The company rewrote the page so each section and each key sentence named the product and plan, and added product names to its table's row labels. A month later, answers attached features to the right products in almost every run.
What do people ask about pronouns and AI extraction?
Does this mean AI cannot understand pronouns?
Models handle pronouns well within a passage. The problem is passages extracted from their context.
Should every paragraph start with the brand name?
No. Start sections and key fact sentences with the subject. Vary the rest naturally.
Does this apply to product descriptions in feeds?
Yes. Feed descriptions are read on their own, so they should name the product in the first sentence.
Published by The Kovmere Team · updated Sep 29, 2026 · More resources