Experiments

Rank in ChatGPT: five experiments, no formulas

Questions about how to rank in ChatGPT arrive daily, and confident answers are cheap. These experiments cost a few hours and leave you with your own data.

AlgoLink research team · Updated 2026-10-28

Start with a confession. We cannot tell you how to rank in ChatGPT, because OpenAI has not published how its search picks sources, and ChatGPT answers vary between runs.

What you can do is observe. ChatGPT search, launched 31 October 2024, shows linked sources, so each answer leaves a trail. Perplexity and Claude with web search do the same, which lets you compare engines on identical prompts.

Treat every run as a sample. Never as a verdict. Use the same wording, a logged-out session where possible, and a date column. Without that discipline the results are anecdotes, however many screenshots you collect.

A word on ethics. Do not spam prompts at an assistant to inflate any count, and do not publish screenshots as proof of a trend. A tidy spreadsheet convinces more people than a folder of images.

The five experiments

Experiment 1: the baseline log

Pick 20 questions your customers ask. Run each in ChatGPT, Perplexity and Google, and record every cited domain. This is your control for everything below.

Experiment 2: wording sensitivity

Rephrase five of the questions three ways each. Count how many cited domains survive. Low overlap tells you single-prompt screenshots mean little.

Experiment 3: change one page

Rewrite one page so the answer opens in two sentences, leave a similar page untouched, and rerun the log after a re-index. One change only, or you learn nothing.

Experiment 4: the neighbourhood check

List the cited domains for one topic and pull their referring domain counts. See whether your site resembles them. The worked pull shows the calls.

Experiment 5: the mention hunt

Search for places that describe your brand without linking to it. Ask for a link or a clearer description, then see whether sightings change over eight weeks.

A log format that stays comparable

Plain CSV is enough. The rows below are made up for illustration and carry no real results.

csv

date,engine,prompt_id,cited_domain,position,ours
2026-10-05,chatgpt_search,q07,example-a.com,1,no
2026-10-05,chatgpt_search,q07,example-b.org,2,no
2026-10-05,perplexity,q07,example-a.com,1,no
2026-10-12,chatgpt_search,q07,example-b.org,1,no

Before you trust a result

The same prompts, word for word, each run
Dates recorded, and the engine's product name noted
At least four weekly runs before comparing
One page changed at a time
A control page left alone
Link data pulled after an index refresh, not before

How to read what comes back

Expect noise. The same prompt can cite different pages an hour apart, and the answer text can change while the sources stay put.

Judge by frequency. Single appearances prove little. If your domain shows in six of eight weekly runs for one question, that is a pattern worth a note. If it shows once, it is a coincidence that happens to flatter you.

Compare engines separately. Google AI Overviews, ChatGPT search, Perplexity and Claude each retrieve in their own way, so a win in one says little about another. The glossary explains why grounding differs between them.

Time budget, for planning. Experiment 1 takes an afternoon. The others are mostly waiting: a re-index can take days or weeks, and nothing in the process can be rushed by paying for it.

What a null result means

If nothing moves, you have learned something real: that your change was not the lever, or that the noise is larger than the effect. Write it down anyway. It will save you from repeating it.