AI visibility proposals often reduce a complex problem to a neat checklist: add FAQs, place a summary block at the top, implement structured data, and wait for citations to climb. Those deliverables are easy to scope and bill. The harder question is whether page formatting actually fixes the reason your content is absent.

Before signing off on template changes, ask the provider to isolate where the breakdown occurs. Does your page directly answer the buyer's question? Does the search system retrieve it? If retrieved, does the model select it as a cited source? A formatting checklist cannot answer those questions on its own.

Why within-website comparisons change the picture

A study by Mahe Chen and Nick Koudas analyzed two commercial sectors: US home internet and niche B2B software. Their most valuable methodological choice was comparing pages within the same website rather than merely across different domains.

Imagine that well-established, frequently cited domains also have author bios, clean schema markup, and summary boxes. If you only compare those domains against smaller, less visible sites, those on-page features can look like the drivers of citation—even when broader domain authority or off-page factors are doing the heavy lifting. Within-site comparisons strip away domain-level advantages, though differences between individual pages still remain.

Once the researchers compared pages on the same domain, the apparent citation advantages of FAQs, summary boxes, author profiles, schema.org markup, and faster load times lost statistical significance. That doesn't mean you should delete helpful summaries or let site speed degrade. It does mean you should be skeptical when an agency attaches a citation guarantee to those elements.

Treat the retrieval finding as a diagnosis, not a budget formula

In the primary ChatGPT analysis—covering 6,178 pages across 362 websites—retrieval metrics explained 32% of the within-site variation in citation counts, while page features explained just 2.8%. You can inspect the full comparison in Section 6.4. Note that these percentages reflect statistical variance explained in an observational dataset, not how you should split your marketing budget.

Methodology matters here. The researchers captured retrieval data using an API replay conducted after the initial 90-day observation window, meaning it did not perfectly reconstruct every original search state. Nor did the study run controlled experiments that actively improved retrieval for randomized pages. In the smaller B2B software sample, the data pointed in the same direction, though the gap between retrieval and page features did not reach statistical significance.

The practical takeaway is straightforward: investigate whether your pages are being discovered before assuming their layout is the bottleneck. However, the paper does not hand you a single technical lever that forces an ignored page into the retrieval set.

Demand evidence at the point of failure

When scoping an audit, pick a handful of commercially critical buyer questions and ask the vendor to show you the full model outputs alongside the underlying source logs. Force the diagnosis to get specific:

  • Question fit: Which exact URL answers the prompt, including specific buyer constraints? A high-level category page may not satisfy a query about a specific integration, regional availability, or edge-case workflow.
  • Observed discovery: What proof exists that the system retrieved the page during its search step? If the provider cannot access retrieval logs, they need to state that plainly. You cannot prove a retrieval failure solely from a missing final citation.
  • Source selection: When the system did retrieve the page, did it cite it? Which competing URLs appeared in the same retrieval run?
  • Brand representation: Did the response mention your brand (linked or unlinked), and did it state your capabilities accurately?

These four checks are practical audit requirements, not experimentally proven fixes from the study. Their job is to ensure any follow-up work targets a verified failure mode.

Make sure your vendor keeps their observation layers honest. An offline content score, a search-enabled API call, and a live consumer web session are three distinct environments. If a consultant uses one as a stand-in for another, ask them to show the validation data supporting that leap.

Turn a broad audit into a targeted project

Different breakdowns require completely different work, even if they all show up on a dashboard as a low visibility score. A page that misses a buyer's constraint calls for targeted content writing. A thorough page that never shows up in retrieval logs raises a discovery question. A page that gets retrieved but misquoted points to an extraction or synthesis issue.

Keep the study's boundaries in mind: it relied on specific API models, English-language US prompts, and a sample limited by unavailable pages, and the authors operated the monitoring platform used to gather the data. You shouldn't blindly extrapolate its exact percentages to your domain. Instead, use its framework to demand a sharper diagnosis—and require local proof on your own target queries before funding sitewide template overhauls.

A credible proposal names the specific URL, documents the missing content or observed pipeline failure, and defines what a successful outcome looks like. That is a much firmer foundation than paying to bolt the same formatting block onto every page.