All content is now TOFU content

Your pricing page is accurate. It lists every tier, every limit, and every add-on. What it doesn't do is explain what a typical customer actually pays, because that depends on how they use the product.

Meanwhile, a competitor has published a comparison suggesting your product gets expensive fast. A prospect asks an answer engine what you cost. The answer takes your limits from your pricing page and its conclusion from your competitor. You're off the shortlist before anyone on your team knows the conversation happened.

Nothing on your page was wrong. Context was missing, and someone else filled the gap.

My colleague Jason McGhee, head of product and engineering at Palmata, used a version of this example in a webinar hosted by Contentful, and it captures a pattern we see often. Content written for customers, developers, and procurement teams, such as pricing pages, technical documentation, support articles, and legal terms, is now being used as evidence when prospects ask their first questions. Bottom-of-funnel content is doing top-of-funnel work. (For the content operations side of this shift, read the companion post on the Contentful blog.)

This post covers why it happens, how to spot it, and how to fix it.

Why answer engines reach for bottom-of-funnel content

Answer engines gather information from across the web, select what seems most relevant and credible for the question, and synthesize it into a single answer. When the question is evaluative (Is this product reliable? What does it cost? What are its limits?), the most specific evidence available is often your most operational content.

That's how a support article ends up in an answer about reliability, or an API reference ends up in a nontechnical comparison. Jason described seeing terms of service, privacy policies, support documentation, and technical docs cited to support answers. Content that once reached only people who had already bought the product is now part of the initial research.

How often this happens, and which sources get chosen, varies by answer engine and can change as models are updated. But the pattern is common enough that every team should check for it.

Accurate, but misleading: Six ways context gets lost

The problem is rarely false information. It's accurate information that loses its context on the way into an answer. That loss usually takes one of six forms:

• Scale: A rare issue reads as a common occurrence. A troubleshooting article for an edge case becomes evidence of a widespread problem.

• Time: Historical information reads as current. An old announcement or retired capability competes with how the product works today.

• Audience: Guidance written for experts lands in a general buyer's answer. Advanced configuration steps can make a product sound more technical or harder to use than it is.

• Conditions: A limit appears without the scenario it applies to. A cap that only affects the highest-volume customers reads as a constraint on everyone.

• Resolution: A known problem appears without its fix. The answer cites the issue, but not the fix that resolved it.

• Comparison: A third party's framing is more explicit than yours. If a competitor or review site states a clear conclusion and your content doesn't address the question directly, their framing is the one that gets used.

That last point is worth stressing. Answer engines build narratives from the evidence available to them, so gaps in your content can shape an answer as much as anything you publish. As I said in the webinar, you can be cited and still be framed in a way that doesn't help you: described as a niche player instead of a leader, or tied to a capability that no longer works the way it used to.

How to tell when it's happening

Visibility metrics alone won't show you this. Knowing you were mentioned or cited doesn't tell you whether you were described accurately or favorably. To see the problem, you need different evidence:

• Mentions vs. citations: whether you're named in an answer, and whether your own content is used as a source.

• AI Perception: whether you’re described positively or negatively in the answer.

• Considered sources: some answer engines, such as ChatGPT, consult a pool of sources but use only some of them in the final answer. If your pages are considered but not cited, they may be close to useful without directly answering the question.

• Adversarial prompts: questions designed to probe your brand's known weak spots.

• Buyer stage: whether your representation changes across discovery, evaluation, and decision questions.

With that evidence, the investigation follows a simple sequence:

1. What questions are prospects asking about your category and your brand?

2. How is your company described in the answers?

3. Which of your pages, and which third-party sources, support that description?

4. Which important pages are retrieved but not used?

5. Does your representation change across buyer stages?

6. Which changes would add missing context or correct an inaccurate narrative?

This is the workflow Palmata is built around. A Palmata report runs this analysis across hundreds of prompts it generates based on adaptive deep research, showing which sources shape how you're represented, and where your own content helps or hurts.

Fixing it without gaming it

Once you know which pages are shaping the wrong narrative, the fix is usually more true context, not technical tricks. Go back to the pricing example. Jason's point was that when a company clearly explains scenario-based pricing, an answer engine can reach a fairer conclusion based on nothing more than accurate information it didn't have before.

That matches something else we've observed. When an answer engine handles a negative question about a brand, it often gives weight to the brand's own content, much as people weigh both sides of a public dispute. In Jason's experience, exploitative tactics seem to be working less and less. That's an observation from our work, but answer engine behavior can change quickly. But it points to the same strategy: clarify, don't manipulate.

Not every fix is the same size. Some are quick wins, like adding an FAQ or clarifying a single paragraph. Others are strategic, such as reworking how a topic is explained across pricing, docs, and support.

Before committing time to either, it helps to know which changes are likely to matter. Palmata models candidate content changes, reruns the relevant scenarios, and compares the resulting answers. If a change doesn't appear to improve the outcome, Palmata doesn't recommend it. Instead, it reworks the changes until it finds the content update that improves your perception in the associated answers. Simulations estimate how answers may change; they can't guarantee how a live answer engine will respond. But they help teams put effort where it's most likely to pay off.

Bring the rest of the company into the report

The pages shaping your AI reputation can come from support, product, legal, and documentation teams, not only marketing. Those teams need to see the evidence so they can reframe their content to match your brand's current story. If your support articles show up as a top-cited source and they're hurting your reputation, your support team should know.

Run your first Palmata report for free to see which sources shape how answer engines describe your brand, and invite your support, product, and legal teams to review it with you.