# How I Use AI | Khadija Zaman

> Where AI sits in how I research and write: the data and the checking are mine, AI drafts, and every claim carries a verified, recalled or inferred label.

Source: https://khadijazaman.com/ai/  ·  Last modified: 2026-10-26

[Home](/) / How I use AI

# How I use AI

I write this site. Given what I work on, you should know exactly where AI fits in.

AI drafts. The data, the point and the checking are mine.

![Khadija Zaman](/static/img/khadija-avatar.webp)**Khadija Zaman** AI Search Manager at Wellows, in search since 2018

## Who writes khadijazaman.com?

Khadija Zaman writes it. She uses AI for first drafts and site code, but the data, the findings, the final wording and every check are hers, and nothing goes live until she approves it. No finding on the site comes from a model, and every published number links to a source you can check.

## Who owns each step

-   Me
-   AI
-   Source

1.  1Source**The data**BigQuery datasets, API probes of AI engines, Search Console
2.  2Me**I find the point**What the data says, and what it doesn't
3.  3AI**AI drafts**Structure and a first pass at the prose
4.  4Me**I check every figure***verified* *recalled* *inferred*
5.  5Me**I publish**Only stats with a source you can check

## The substance is mine

Every piece here starts from something I measured. That's a BigQuery dataset, a probe I sent to an AI engine's API, Search Console data, or something I watched happen on a product I work on. AI never supplies the finding. If I can't point to where a number came from, it doesn't go in.

The largest example is my [cross-engine AI citation study](https://wellows.com/blog/ai-citation-overlap-study/) on the Wellows blog. The raw pool was 22,749,707 citations across 1,146,483 questions, drawn from Wellows' internal citation dataset. A model can help me phrase what that data shows. It can't produce the data, and it can't tell me whether the finding is real.

## What AI does here, task by task

"I use AI" covers everything from spellcheck to fabricated studies, so here is the split on this site.

| Task | What AI does | What I do |
| --- | --- | --- |
| Articles and studies | Structure and a first pass at the prose | The finding, the argument, every number, the final cut |
| Site code | Drafts changes to templates, scripts and styles | Review each change and merge it; nothing reaches the live site any other way |
| Research data | Nothing. The datasets come from BigQuery, engine APIs and Search Console | Pull, filter and analyse it |
| Checking | Nothing it can be trusted with alone | Six verification passes on anything with numbers in it |
| Answers to visitors and agents | Nothing. The answers endpoint has no model in the loop | Write the two-sentence answer on each page that it quotes verbatim |

That last row is deliberate. When an agent asks this site a question, the endpoint matches it against the direct answers I published and returns one word for word, with its source page. If nothing matches well enough, it returns no answer at all rather than a plausible guess.

## Every claim gets a label

Before anything goes live, I sort its claims into three kinds. **Verified** means I checked it against a source you can check too. **Recalled** means it's from memory, and I say so. **Inferred** means it's my reading of the data, not the data itself.

The [R2A framework](/frameworks/r2a/) applies the same idea to recommendations: each one carries one of five evidence grades, so you can see how sure I am before you act on it.

## Fast drafting made the checking harder, not optional

The citation study went live on 29 July 2026. By 3 August I had rewritten it twelve times, and the verification behind those rewrites took three weeks. Drafting was never the slow part.

The checking found things I wasn't looking for. The post said every question was collected in the United States; the data said 84.06% United States across 27 markets. A permutation-null test then knocked out two headline findings, and I [published the revision](/research/citation-study-retest/) beside the original numbers. Three weeks of numeric passes also missed a bolded sentence that contradicted the paragraph above it. A straight read-through as a reader caught it.

That last one is the reason I don't let a model check its own draft. A draft that reads smoothly is exactly the kind that hides a contradiction.

## Using AI is part of the research

I study how AI engines decide which brands and pages to cite. Drafting with the same kind of models shows me how they summarise, what they drop and where they fill gaps with something plausible. When a draft gets a fact wrong, that's a data point, not just an edit.

The [BERAP Map](/frameworks/r2a/berap/) turns that habit into an instrument. It sends repeated prompts to AI engines and grades each answer against a dated table of facts about a brand, so recall gets measured against an answer key instead of a feeling.

## Why speed matters

AI search moves fast. Engines change models and retrieval, and a finding about how one of them cites can go stale quickly. If writing up a result takes weeks, the result can be out of date before it's live. AI shortens the write-up. It doesn't shorten the checking.

## What Google says about it

Google's position, from its [February 2023 guidance on AI-generated content](https://developers.google.com/search/blog/2023/02/google-search-and-ai-content), is that it rewards helpful content however it's produced. What breaks its spam policies is using automation mainly to manipulate rankings. I don't write for rankings anyway. My test is whether you can use the page today.

## If that's not for you

That's fair. Now you know exactly where AI sits in how I work, and you can decide whether this site is for you. Every number here should have a source you can follow. If you find one that doesn't, [tell me](/contact/) and I'll fix it.

---
Markdown twin of https://khadijazaman.com/ai/, generated from the rendered page at build time. Cite the HTML URL. How to cite: https://khadijazaman.com/llms.txt