---
title: "Did the Fix Get You Cited? A Six-Week Test for AI Overviews, ChatGPT and Perplexity"
description: "A checklist tells you what to fix. This six-week protocol tests which landing-page changes earn citations on your own queries—and whether a tool does more than report them."
url: "https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai"
image: "https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/ae7fa7d6-6913-4c5a-9cff-6069bc752175.png"
published: "2026-11-02T06:02:35.702Z"
modified: "2026-11-02T06:02:35.762Z"
---

November 2, 2026 · 11 min read

# Did the Fix Get You Cited? A Six-Week Test for AI Overviews, ChatGPT and Perplexity

[Alexander PerelmanHead Of Product @ groas](https://groas.com/author/alexander-perelman)[LinkedIn](https://www.linkedin.com/in/alexander-433793253/)

![Two rows of tiny model shops, one freshly repainted and one left alone, with a tiny lab-coated tester watching spotlights land unpredictably across both.](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/ae7fa7d6-6913-4c5a-9cff-6069bc752175.png)

In this article

1. [Define the outcome before you edit](#define-the-outcome-before-you-edit)
2. [Run four intervention rounds, one family per week](#run-four-intervention-rounds-one-family-per-week)
3. [Read fetches, citations and visits in that order](#read-fetches-citations-and-visits-in-that-order)
4. [Set the win condition before you see the score](#set-the-win-condition-before-you-see-the-score)
5. [Put software through the same control-group test](#put-software-through-the-same-control-group-test)

You fixed the landing page. On Tuesday, Google cites it; on Thursday, it cites a competitor; on Friday, it cites nobody. That is why I am taking a second look at [_The AI Overview Pre-Flight: 22 Landing Page Checks Before You Blame Google_](https://groas.com/post/the-ai-overview-landing-page-pre-flight-2): a fixed page is not the same as a cited one, and AI answers shift week to week.

A checklist tells you what to change. **Only a controlled before-and-after test on your own target queries tells you which changes earn citations.** The question here is narrow: which page changes increase how often your target URL gets cited for queries that matter to your business? Use the same prompts and cadence, compare treatment pages with controls, and score [citation rate per query per surface rather than rank position](http://ziptie.ai/blog/how-ai-search-tracking-actually-works/).

The schedule is six weeks: two for the untouched baseline, then four one-week intervention rounds. If a change does not beat its controls in that window, you have not shown that it worked. You have not proved that it never will.

## Define the outcome before you edit

I used to treat a ranking fix and a citation fix as the same job. I was wrong. One [analysis reports that 38% of AI Overview cited pages rank in Google’s top 10 for the query](https://adamarant.com/en/blog/how-to-get-cited-in-google-ai-overviews-and-chatgpt-in-2026), down from 76% in July 2025; it puts the comparable figure for ChatGPT cited links at about 12%. Retrieval goes past page one.

This test measures whether your URL gets cited on your queries. It cannot tell you why Google changed an answer template last Tuesday, stop a competitor from rewriting a page, or produce a universal ranking factor. AI answers are [non-deterministic: the same query can cite different sources across sessions](http://ziptie.ai/blog/how-ai-search-tracking-actually-works/). That is the reason to repeat prompts and keep controls, not a reason to abandon the test.

**Decide what counts as a citation before you collect the first answer.** A brand mention is not a citation to your target URL. A citation to a third-party review is not one either. Keep those observations, but do not score them as page wins.

### 1. Choose 10 to 20 buying queries

Start with prompts a buyer would actually use, not your brand name: “How much does emergency drain cleaning cost in Austin?” or “Best payroll software for a 20-person agency.” Put 10 to 20 prompts in a sheet and assign each to the page that should own it. A service page owns the service question; a pricing page owns cost. If two pages could answer the same prompt, pick the target URL now. Otherwise, the result will read as noise.

Other practitioners use a [10-to-20-prompt baseline, tracking mention rate and cited URLs](https://medium.com/@arevalopeachy/how-to-build-an-ai-visibility-system-that-actually-works-2c2784314182). For this test, keep both fields, but score the **target URL’s citation rate**. That is the yardstick for the next six weeks.

### 2. Pair treatment pages with controls

Pick four to six comparable pages to treat and a matched set to leave alone. Say a home services site has ten city pages: treat five and hold five as controls. Look for similar templates, traffic and age. Record each pair before the baseline starts.

Do not rewrite every treatment page in every round. Assign pages to intervention families in advance, then change only the assigned treatment pages when their round arrives. Leave their controls untouched. Compare each page’s post-change runs with its own pre-change runs and with its paired control. With a small page set, one weak pair can swing the result; write that limitation in the notebook rather than hiding it in a chart.

Controls give you a check on changes outside your pages. If treatment citations rise while controls stay flat, the edit has a stronger case. If both rise together, the edit does not get credit. **A page is not its own control when answers swing from session to session.**

### 3. Collect a two-week citation baseline

Run every prompt on Google AI Overviews, ChatGPT with browsing, and Perplexity. Keep the wording, logged-in state and time of day consistent. Run sessions three times a week for two weeks; run each prompt three times per session. Save the full answer and every cited URL, not just a yes-or-no mark.

Score each surface separately. In one week, nine runs give you a simple rate: how many cited the target URL? Use the two baseline weeks to see how much that rate moves before you touch the page, then compare each intervention week against the baseline and its matched control. Google, ChatGPT and Perplexity [share only 10% to 15% citation overlap](http://ziptie.ai/blog/how-ai-search-tracking-actually-works/), so a blended score can hide the only surface where an edit helped.

### 4. Check whether AI fetchers can reach the pages

Before rewriting copy, pull two weeks of server logs and search for `GPTBot|OAI-SearchBot|ChatGPT-User|PerplexityBot|Perplexity-User|ClaudeBot|Claude-SearchBot`. Count the top paths each agent fetched, split results by status code, and separate treatment from control pages. In particular, [check live fetches for 403s and 429s](https://dev.to/furqank729/your-access-logs-already-show-which-pages-ai-assistants-are-reading-1di5). A blocked user-triggered fetch can look like a writing problem in a citation report. It is an access problem.

The [agent names describe different jobs](https://forgemesh.io/blog/bots-that-read-your-website):

- **Training crawlers:** `GPTBot` and `ClaudeBot`.
- **Search indexers:** `OAI-SearchBot` and `PerplexityBot`.
- **User-triggered fetchers:** `ChatGPT-User` and `Perplexity-User`.

![Server-log lines for AI fetchers beside a sheet of citation scores](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/277e88b2-f28b-42ac-aea2-eac5b9b957f3.png)

Do not trust a user-agent string by itself. [Spoofed bot claims are common; verify by reverse DNS or vendor IP ranges](https://forgemesh.io/blog/bots-that-read-your-website) before allowlisting an agent or declaring that one is blocked. For Google AI Overviews, [the page must be indexed and eligible for a Search snippet](https://adamarant.com/en/blog/how-to-get-cited-in-google-ai-overviews-and-chatgpt-in-2026). Check URL Inspection, canonical and rendered HTML. For ChatGPT Search, [`OAI-SearchBot` governs search visibility while `GPTBot` handles training](https://adamarant.com/en/blog/how-to-get-cited-in-google-ai-overviews-and-chatgpt-in-2026); the robots file can treat them differently. For Perplexity, [check `PerplexityBot` access and whether the CDN returns 200](https://dev.to/furqank729/your-access-logs-already-show-which-pages-ai-assistants-are-reading-1di5). Record the access state now. Otherwise, a blocked page can make a later copy test meaningless.

## Run four intervention rounds, one family per week

**The rule is one assigned family, one week, one comparison with unchanged controls.** Do not give the same treatment page a new answer, new evidence and a new date at once, then claim you know which helped. Record the exact edit and publication time. Keep measuring all queries on the original cadence, including those for pages you have not changed.

### Week 3: Fix fetchability

If assigned treatment pages show access failures, fix those first: indexing and snippet eligibility, canonical, answer in rendered HTML, and a CDN response of 200 for the verified bot you need. Do not use this round to rewrite the pitch. Recheck the logs after the fix.

The mechanism comes before the prediction: an answer system cannot use a passage it cannot retrieve. **I expect successful fetches to appear before any reliable citation lift.** If access still fails, do not interpret flat citations as a verdict on the copy. If access was already clean, record that and do not manufacture a fetchability change just to fill the week.

### Week 4: Move the answer up

On the assigned treatment pages, put the buyer question in an H2 and answer it in 40 to 60 words immediately below. Make the passage self-contained: price, turnaround, coverage area or minimums where those facts belong. Leave the matched controls as they are.

The mechanism is extraction. A system looking for an answer has less assembly to do when the relevant passage states it plainly near the top. I used to bury answers under brand copy because the introduction read better. A reader looking for the price did not need my introduction. **Measure whether the target URL gets cited more often, not whether the new opening sounds more polished.**

### Week 5: Make the evidence quotable

On this round’s treatment pages, replace vague claims with accurate first-party facts the page can support: a current price and date, named service areas, model numbers, turnaround in days or license numbers. Put the business or product name and relevant location near the fact instead of expecting a header to supply the context.

The mechanism is extraction confidence. A sentence that names the fact, its owner and its scope can stand alone; a sentence that says “it starts low” cannot. If a page says _affordable plans starting low_, the answer still has to guess what “low” means. A specific, current price gives it something it can cite. **Only publish facts you can keep correct.** Watch the cited URL and the fact the answer carries, not just the mention count.

### Week 6: Show a real update

Give this round’s treatment pages a truthful updated date and a short changelog: what changed, when, and which price or coverage details are current. Do not stamp today’s date on an unchanged page. The mechanism is freshness with visible proof, not a decorative timestamp. Pages not updated quarterly were [three times more likely to lose citations in AirOps data summarized in this visibility-system breakdown](https://medium.com/@arevalopeachy/how-to-build-an-ai-visibility-system-that-actually-works-2c2784314182).

A real change gives a fetcher something new to find. A new date without new information gives you a calendar trick to measure. **Keep the controls untouched and record the actual change.**

## Read fetches, citations and visits in that order

The primary metric is **citation rate per query per surface**: in nine runs during an intervention week, how many times did the target URL appear as a citation? Keep separate columns for Google AI Overviews, ChatGPT and Perplexity. Compare each result with that query’s two-week baseline and its paired control.

The secondary metric is successful **AI fetches per page**, before and after the edit. Logs can show whether an agent requested a page; they do not prove that the answer used the retrieved passage. A newly successful fetch with citations still flat is useful, but it is not a citation win. If verified fetches still fail, return to access before rewriting sentences again.

Then follow the [_Crawled > Cited > Clicked_ sequence](http://ziptie.ai/blog/how-ai-search-tracking-actually-works/):

- **Not fetched:** Check the access problem before judging the content.
- **Fetched but not cited:** Inspect answer placement, evidence and the URLs that won instead.
- **Cited but not clicked:** Record the citation win separately from traffic. That tracking breakdown notes that 93% of cited answers are zero-click.

A mention is not pipeline. Tag cited URLs and watch assisted visits, form fills and booked jobs during the test. A page that rises from one citation in nine Perplexity runs to five and sends qualified calls has done more useful work than one that gets cited with the wrong price. **Ship citations that carry the right fact to the right page, not citation count alone.**

## Set the win condition before you see the score

Noise is the default explanation. My proposed bar for this small test is a gain of at least three citations out of nine runs on the same surface, while the paired control moves by no more than one. A move from two-in-nine to five-in-nine on ChatGPT, with its control stuck at one-in-nine, is worth following up. Two-in-nine to three-in-nine while the control also climbs is not a win for the edit.

That bar is a decision rule, not a statistical proof. Practitioners recommend [four-to-eight-week cycles rather than judging a single week’s AI visibility](https://medium.com/@arevalopeachy/how-to-build-an-ai-visibility-system-that-actually-works-2c2784314182). This protocol uses two baseline weeks, four intervention weeks and controls to make a short test more informative. If results are close, keep measuring instead of rounding a near miss into a case study.

Expect split decisions across surfaces. A fetchability change may show up in Perplexity logs before its citation rate moves; a clearer answer may get picked up on one surface and ignored on another. When treatment and controls move together, write down a shared movement, not your win. A result with successful fetches but flat citations says access was not the whole problem. A result with citations but no matching fetch in your logs calls for a check of the cited URL before you credit the edited page. **The next action follows the observed failure, not the round you hoped would work.**

## Put software through the same control-group test

A monitoring dashboard can tell you that citations moved. It cannot claim credit for moving them unless something it did changed the page or its access. The [Semrush AI Visibility Toolkit, listed at $99 per month as an add-on in this comparison](https://www.mersel.ai/blog/mersel-ai-vs-semrush-aio-feature-breakdown), tracks mentions and share of voice across ChatGPT, Gemini and AI Overviews; the comparison says content creation, schema and publishing remain with your team. What the deck calls visibility, I call a chart of work someone still has to do. The distinction between [tools that help pages get featured and tools that monitor mentions](https://groas.com/post/tools-to-get-featured-in-google-ai-overviews) matters most when you run the test.

![Monitoring dashboard beside an editor changing a live page](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/cf30e1ed-21f5-4354-8b19-8b42323267fa.png)

Treat a vendor trial as another intervention round. Baseline its assigned pages, leave matched controls alone, and agree on three answers before the vendor touches anything:

1. **Which intervention family will it ship:** fetchability, answer placement, evidence or freshness?
2. **Who publishes the change:** the tool, your developer or nobody?
3. **Which fetch and citation outcome will it check afterward?**

If the vendor cannot name the intervention, you are buying monitoring. groas is built to take on the page changes and technical fixes, log its actions and reasoning, and charge a flat fee rather than hand your team a report and a queue of tickets. It still has to face the same test: treatment pages against controls, on your queries. The longer version of that control-group logic is in [_Did Your Page Change Earn an AI Citation? Run a Control-Group Test_](https://groas.com/post/did-that-page-change-get-you-cited-or-di).

At the end of six weeks, carry a winning change to comparable pages and keep measuring. Leave a null result out of the rollout; investigate a failed fetch before paying anyone to rewrite copy. **If treatment citations climb and controls stay flat, you have a reason to keep the change. If they do not, you have a reason to change the plan.**

## Frequently Asked Questions

### What must be decided before collecting citation data?

You must define what counts as a citation to your target URL before starting the test. A brand mention or a link to a third-party review should be recorded but should not be scored as a page win.

### How should landing page changes be tested against controls?

Test pages must be paired with comparable controls that are left untouched during the intervention. A page is not its own control when AI answers swing from session to session, requiring a matched set for accurate comparison.

### What server log checks should be performed before rewriting website copy?

Pull two weeks of server logs to search for specific AI bot user-agent strings such as GPTBot and PerplexityBot. Checking live fetches for 403s and 429s verifies if a blocked fetch is an access problem instead of a writing problem.

### In what order should landing page changes be implemented during the intervention rounds?

The recommended sequence is to fix fetchability issues first, then move the answer up using an H2 and a short paragraph. Next, make the evidence quotable, and finally, show a real update with a current date.

### What is the primary measurement for determining if a page change succeeded?

The primary metric is the citation rate per query per surface, which counts how many times the target URL appears as a citation during a set number of runs. The secondary measurement is the successful AI fetches per page before and after the edit.

### What specific criteria defines a successful change in the six-week test?

The proposed win condition is achieving a gain of at least three citations out of nine runs on the same surface. This must occur while the paired control moves by no more than one citation.

### How should one evaluate AI visibility tools compared to this testing protocol?

Tools monitor mentions and share of voice across different AI services. While these tools track visibility metrics, a control-group test is required to prove that a specific page change actually earned the citation.

## Paid Ads & SEO, Run For You

Businesses pay one flat monthly fee per service, agencies resell it under their own brand, and groas does the work either way.

[See Pricing](https://groas.com/pricing)

## Structured data

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://graph.groas.com/entity/groas","name":"groas","url":"https://groas.com/","description":"groas is a done-for-you marketing service, not software: elite people working with proprietary technology systems run search ads, Meta ads, SEO and AI search for businesses, and white label for agencies, with a dedicated senior strategist owning every account.","logo":{"@type":"ImageObject","url":"https://groas.com/icon-512.png","width":512,"height":512},"sameAs":["https://www.linkedin.com/company/groas/"]},{"@type":"WebSite","@id":"https://groas.com/#website","name":"groas","url":"https://groas.com/","inLanguage":"en-US","publisher":{"@id":"https://graph.groas.com/entity/groas"}},{"@type":"WebPage","@id":"https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai#webpage","url":"https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai","name":"Did the Fix Get You Cited? A Six-Week Test for AI Overviews, ChatGPT and Perplexity","description":"A checklist tells you what to fix. This six-week protocol tests which landing-page changes earn citations on your own queries—and whether a tool does more than report them.","inLanguage":"en-US","isPartOf":{"@id":"https://groas.com/#website"},"about":{"@id":"https://graph.groas.com/entity/groas"},"primaryImageOfPage":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/ae7fa7d6-6913-4c5a-9cff-6069bc752175.png"},"breadcrumb":{"@id":"https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai#breadcrumb"},"dateModified":"2026-11-02T06:02:35.762Z","mainEntity":{"@id":"https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai#article"}},{"@type":"BreadcrumbList","@id":"https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://groas.com/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://groas.com/blog"},{"@type":"ListItem","position":3,"name":"Did the Fix Get You Cited? A Six-Week Test for AI Overviews, ChatGPT and Perplexity","item":"https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai"}]},{"@type":"BlogPosting","@id":"https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai#article","url":"https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai","mainEntityOfPage":{"@id":"https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai#webpage"},"isPartOf":{"@id":"https://groas.com/blog#blog"},"headline":"Did the Fix Get You Cited? A Six-Week Test for AI Overviews, ChatGPT and Perplexity","description":"A checklist tells you what to fix. This six-week protocol tests which landing-page changes earn citations on your own queries—and whether a tool does more than report them.","datePublished":"2026-11-02T06:02:35.702Z","dateModified":"2026-11-02T06:02:35.762Z","image":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/ae7fa7d6-6913-4c5a-9cff-6069bc752175.png"},"author":{"@id":"https://groas.com/author/alexander-perelman#person"},"publisher":{"@id":"https://graph.groas.com/entity/groas"},"keywords":"page, cited, pages, citation, fix, test, answer, before","wordCount":2297,"inLanguage":"en-US"},{"@type":"Person","@id":"https://groas.com/author/alexander-perelman#person","name":"Alexander Perelman","jobTitle":"Head Of Product @ groas","description":"My background includes Goldman Sachs and Stanford Computer Science.","image":"https://groas.com/media/blog/4f17cac3a81acc224cc1b2cfab8f94bbc295a3089d13e8421125771ef8d4064f.jpg","worksFor":{"@id":"https://graph.groas.com/entity/groas"},"sameAs":["https://www.linkedin.com/in/alexander-433793253/"]},{"@type":"FAQPage","@id":"https://groas.com/post/did-the-fix-get-you-quoted-a-four-week-test-for-earning-ai#faq","mainEntity":[{"@type":"Question","name":"What must be decided before collecting citation data?","acceptedAnswer":{"@type":"Answer","text":"You must define what counts as a citation to your target URL before starting the test. A brand mention or a link to a third-party review should be recorded but should not be scored as a page win."}},{"@type":"Question","name":"How should landing page changes be tested against controls?","acceptedAnswer":{"@type":"Answer","text":"Test pages must be paired with comparable controls that are left untouched during the intervention. A page is not its own control when AI answers swing from session to session, requiring a matched set for accurate comparison."}},{"@type":"Question","name":"What server log checks should be performed before rewriting website copy?","acceptedAnswer":{"@type":"Answer","text":"Pull two weeks of server logs to search for specific AI bot user-agent strings such as GPTBot and PerplexityBot. Checking live fetches for 403s and 429s verifies if a blocked fetch is an access problem instead of a writing problem."}},{"@type":"Question","name":"In what order should landing page changes be implemented during the intervention rounds?","acceptedAnswer":{"@type":"Answer","text":"The recommended sequence is to fix fetchability issues first, then move the answer up using an H2 and a short paragraph. Next, make the evidence quotable, and finally, show a real update with a current date."}},{"@type":"Question","name":"What is the primary measurement for determining if a page change succeeded?","acceptedAnswer":{"@type":"Answer","text":"The primary metric is the citation rate per query per surface, which counts how many times the target URL appears as a citation during a set number of runs. The secondary measurement is the successful AI fetches per page before and after the edit."}},{"@type":"Question","name":"What specific criteria defines a successful change in the six-week test?","acceptedAnswer":{"@type":"Answer","text":"The proposed win condition is achieving a gain of at least three citations out of nine runs on the same surface. This must occur while the paired control moves by no more than one citation."}},{"@type":"Question","name":"How should one evaluate AI visibility tools compared to this testing protocol?","acceptedAnswer":{"@type":"Answer","text":"Tools monitor mentions and share of voice across different AI services. While these tools track visibility metrics, a control-group test is required to prove that a specific page change actually earned the citation."}}]}]}
```