Position-Adjusted Word Count (PAWC): definition and limits

Position-adjusted word count is the primary visibility metric in GEO: Generative Engine Optimization (Aggarwal et al., arXiv 2311.09735, published at KDD 2024). For each source in a generated answer, it counts the words attributed to that source. It then adjusts the count by position, so text that appears earlier in the answer weighs more than text near the end.

The result is a share-of-answer number for one source in one response. Researchers compare it before and after editing the source to estimate whether the edit made the source more visible. For a full walk-through of the study, see our digest of the original GEO paper.

The GEO paper's setup has three steps:

1. The authors pick a query and a fixed set of sources.
2. They place those sources directly into the model's context and generate an answer.
3. They edit one source (for example, adding quotations or statistics), regenerate, and compare PAWC.

No live index is queried. Every source is already in front of the model, so the experiment measures generation and citation, not retrieval.

The circularity follows from the edits. Several tested methods add text to the source. If the model reproduces that added text, the source's attributed word count goes up. Part of the measured gain is therefore the intervention's own words appearing in the answer. That is a real effect inside the lab setup, but it is not independent evidence that the page became more useful or more findable.

One result cuts the other way: the paper found keyword stuffing decreased PAWC (17.8 against a 19.5 baseline).

PAWC sits at the last of four stages. A brand has to clear parametric knowledge or access, then retrieval, before generation even starts. PAWC only measures that final stage, on documents already chosen.

That is why its findings transfer poorly. The 2026 critical survey of 45 GEO studies (arXiv 2607.14035) concluded that already-retrieved content can causally alter its citation or use, but no reviewed technique shows a stable, longitudinal, cross-platform effect on discoverability. The C-SEO Bench study went further and found most conversational SEO methods largely ineffective, and often harmful to ranking in context, while position bias dominated.

The next move is to measure what a real engine does with your brand, per engine and per question. Our playbook on AI visibility KPIs you can defend covers which numbers hold up.

"GEO gets you up to 40% more visibility"

The figure traces to PAWC gains on the authors' own engine, with sources pre-loaded. Even the one test labelled "in the wild" uploaded source text to Perplexity as files, because the authors noted perplexity.ai does not let users specify source URLs. The paper reports a 30 to 40% band across three methods. It never reports the itemised +41, +40 and +30 figures in circulation.

"PAWC is a metric I can track for my site"

It isn't. You can't see the context window of ChatGPT or Google, and no engine exposes per-source word attribution. There's also no answer position to read, which our entry on average position explains.

"Longer pages score higher, so write more"

PAWC counts words in the answer, not on the page. Google says there's no ideal page length, covered in ideal content length for AI search.

• Position bias: the tendency for documents placed earlier in context to be cited more.
• FActScore: a factuality metric that scores supported atomic facts, a different question from visibility.
• Citation rate vs citation share: two real-world citation metrics and how to read them.
• Average position: why rank-style numbers don't exist for AI answers.