ChatGPT SEO prompts: 28 copy-paste prompts, each with a check.

Below are [N] customer emails, sales call notes and support
tickets. List the problems these customers describe, written as
short phrases a person might type into Google. Group the phrases
by problem and quote the line each one came from.

Do not add a phrase that no quote supports. Do not estimate
search volume.

NOTES:
[PASTE]

Check: Every phrase has a quote behind it. Then run the list through a keyword tool and drop anything with no measured demand.

My seed keyword is [SEED]. My site sells [PRODUCT] to
[AUDIENCE]. List 40 searches a buyer might type on the way to
buying, grouped as: questions (how, what, why, can), problems,
tasks, comparisons (vs, alternatives) and best-of lists. Mark
each one informational, commercial or transactional.

Do not give search volume, difficulty or traffic estimates.

Check: Treat the list as guesses to test, not data. Look each search up in Search Console or a keyword tool before it goes on a plan.

Below are the titles and URLs of the top 10 Google results
for [KEYWORD], copied from a logged-out search in [COUNTRY].

1. For each result, give the page type: guide, list, product
   page, category page, tool, forum thread or video.
2. Say which page type holds most of the top 10.
3. Name the result closest to what I should build, and why.

If no page type holds at least four of the 10, say the intent
is mixed and stop.

RESULTS:
[PASTE]

Check: The page you plan matches the type that holds most of the top 10. If it matches none of them, pick a different keyword.

The attached CSV has one row per keyword with columns:
keyword, volume, difficulty, and url_1 to url_10 (the top 10
ranking URLs). Using data analysis:

1. Put two keywords in the same cluster when they share at
   least 4 of their top 10 URLs.
2. For each cluster give the head keyword (highest volume),
   the total volume and the member keywords.
3. List the keywords that joined no cluster.

Do not change, round or add any volume. Show the code and the
number of rows you read.

Check: The row count matches the file. Search two keywords from one cluster: if Google shows them different pages, the threshold is too loose.

The attached file is a Search Console performance export
with queries, pages, clicks, impressions and average position.
Using data analysis, list queries with at least [N] impressions
where the best ranking page sits below position 20. Group them
by topic. For each group, say whether an existing page should
be expanded (name it) or a new page is needed.

Show the code and the number of rows you read.

Check: An export from the Search Console interface stops at 1,000 rows, so long-tail queries are missing. Treat the output as the top of the list, not all of it.

Below are the H2 and H3 headings of the five pages ranking
for [KEYWORD], each block labelled with its URL.

1. List subtopics at least three of the five cover.
2. List subtopics only one covers.
3. List questions a reader would still have that none answers.

Quote the heading behind every item in lists 1 and 2.

HEADINGS:
[PASTE]

Check: Every item in lists 1 and 2 points at a real heading. List 3 is where your page can add something, and also where the model guesses, so read it critically.

Write the 40 to 60 word passage that opens a page answering
[QUERY]. Start with the answer itself, not background. Use only
the facts in FACTS. After the passage, list any fact it needed
that FACTS does not contain.

FACTS:
[PASTE YOUR NOTES AND SOURCES]

Check: Read the passage on its own. It should still make sense lifted out of the page, which is how an answer engine would quote it.

Interview me about [TOPIC] so the article contains things the
ranking pages cannot. Ask one question at a time, eight at
most: what we have tested, numbers from our own data, mistakes
we made, and views we hold that others in [INDUSTRY] do not.

After my last answer, turn my answers into bullet points for a
content brief. Use my words. Do not add facts I did not give.

Check: At least three points in the brief come from your answers, not from the ranking pages. If none do, the page has nothing the top five lack.

Using the outline gaps, the opening passage and the interview
points above, write a content brief with these fields:

- target query and search intent
- page type
- opening passage (as written above)
- H2 outline, with one line on what each section must show
- three internal pages to link to, from this list: [URL LIST]
- every fact still needing a source, marked [SOURCE NEEDED]
- what to leave out, and why

Do not set a word count.

Check: No [SOURCE NEEDED] survives into drafting. The brief sets scope, not length: Google asks whether you write to a word count because you heard it prefers one.

Below are the People Also Ask questions and related searches
Google shows for [KEYWORD]. Sort each one into:

A. already answered by the outline
B. needs a sentence in an existing section (say which)
C. needs its own H2
D. belongs on a different page (say what page)

OUTLINE:
[PASTE]

QUESTIONS:
[PASTE]

Check: Put the answers where readers need them. Google stopped showing FAQ rich results on May 7, 2026, so a question block built for markup earns nothing extra in search.

Page URL: [URL]
H1: [H1]
Target query: [QUERY]
What the page delivers, in one line: [PROMISE]

Write 10 title tags. Five start with the target query. Five
lead with a specific number or outcome the page contains. Aim
for 50 to 60 characters. No year unless the page is dated. No
word or number the page cannot back up.

Check: Count the characters yourself, because models miscount them, and make sure the title says what the H1 says. Google builds title links from several sources, headings and anchor text included, not only the title tag.

Write five meta descriptions for the page below, each 120 to
155 characters. Each one says what the reader gets, includes
one specific detail from the page (a number, a step, a named
example) and uses [QUERY] or a close variant once. Do not
repeat the title.

PAGE:
[PASTE]

Check: Count characters outside the chat. Google builds snippets mostly from the page itself and uses the meta description when it describes the page more accurately, so accuracy beats keywords.

Below is my page's heading outline (H1 to H3) and the query it
targets: [QUERY]. Flag:

- headings that do not say what their section answers
- sections that answer a different query
- skipped heading levels
- where the main answer to [QUERY] currently sits

Suggest rewrites for flagged headings only.

OUTLINE:
[PASTE]

Check: The main answer belongs in the first section under the H1. If the model finds it further down, move it before you touch anything else.

Rewrite each paragraph below so its first sentence answers the
question its heading asks, in plain words, and the rest of the
paragraph supports it. Keep every fact, number and name exactly
as written. Do not add facts. Mark any section whose heading
question the text never answers.

SECTIONS:
[PASTE]

Check: Compare the rewrite with the original line by line. Any number or name that was not there before is an invention, not an edit.

I am uploading [N] images from the page below. For each one,
write alt text in one plain sentence that says what the image
shows and why it is on this page. Use the target query only if
the image actually shows that thing.

PAGE:
[PASTE]

Check: Read each line with the image hidden: it should tell you what you would see. Google's image guidance warns that filling alt attributes with keywords is a bad experience and can get a site seen as spam.

My page is a [PAGE TYPE] that contains: [WHAT IS VISIBLE ON
THE PAGE]. List the schema.org types that describe it
accurately. For each type, say whether you believe Google still
shows a rich result for it, and which required properties my
page can fill from its visible content. Write NOT SURE instead
of guessing about Google support.

Check: Compare the answer with Google's structured data search gallery yourself. FAQ and HowTo are no longer in it, whatever the model remembers.

Write JSON-LD for the page below using the [TYPE] type. Use
only values that appear in the page text I pasted: headline,
author, dates, prices, ratings, steps. If a required property
has no visible value, leave it out and list it under MISSING.
Do not invent ratings, review counts or prices.

PAGE TEXT:
[PASTE]

Check: Find every value on the live page with Ctrl+F. Google's structured data guidelines say not to mark up content that is not visible to readers.

Below is my JSON-LD and the errors and warnings the Rich
Results Test reported for it. Fix only what the errors name and
explain each change in one line. Do not add a property to clear
a warning unless its value is visible on the page.

JSON-LD:
[PASTE]

TEST OUTPUT:
[PASTE]

Check: Re-run the Rich Results Test, then the URL Inspection tool on the live URL. A clean answer in the chat window is not a test.

My site name is [NAME]. Homepage: [URL]. Logo: [URL]. Social
profiles: [URLS]. The main sections and their URLs: [LIST].

1. Write one Organization JSON-LD block for the homepage.
2. Write a BreadcrumbList template for pages in [SECTION],
   with positions and full URLs.

Use only the URLs I gave you.

Check: Every URL in the output is one you supplied, and the breadcrumb trail matches the one readers see on the page.

Here is my list of URLs with page titles:
[PASTE LIST]

Here is a new article:
[PASTE ARTICLE]

Suggest up to six internal links from the article. For each,
quote the exact sentence, the words to use as anchor text and
the target URL. Use only URLs from my list. Skip a link rather
than force one.

Check: Every target is on your list and loads. A URL the model wrote from memory is the first thing to delete.

Below are the internal links on my page: anchor text, target
URL and the target page's title. Flag:

- generic anchors (click here, read more, this post)
- anchors that describe a different topic from the target
- one anchor used for two different targets

For each flag, suggest a replacement built from words already
in the sentence.

LINKS:
[PASTE]

Check: Read each anchor on its own. Google's link guidance says good anchor text is descriptive, reasonably concise and relevant to both pages, and suggests exactly that out-of-context test.

The attached files are a crawl export of my internal links
(source URL, target URL, anchor text) and my sitemap URL list.
Using data analysis:

1. List sitemap URLs with zero internal links pointing to them.
2. List sitemap URLs with exactly one.
3. For each, suggest three crawled pages on the closest topic,
   matched on title words, as candidate link sources.

Show the code and the row counts of both files.

Check: Fix the zero-link list first. Google says every page you care about should have a link from at least one other page on your site.

Here are the [N] pages in my [TOPIC] cluster, with title
and URL:
[PASTE]

1. Outline a hub page that links to each of them once, with a
   one-sentence reason to click.
2. For each cluster page, name the two or three sibling pages
   it should link to.
3. List topics the cluster is missing, separately.

Use only pages from my list in parts 1 and 2.

Check: Every page in the plan links back to the hub, and every URL came from your list.

The attached CSV is a crawl of my site with columns: URL,
status code, indexability, title, meta description, H1,
canonical. Using data analysis, list:

- pages that return anything other than 200
- pages whose canonical points to another URL
- duplicate titles and duplicate H1s
- missing titles, H1s or meta descriptions

Group by issue with page counts, sorted by count. Show the code
and the number of rows you read.

Check: Open five flagged URLs yourself. One misread column turns a whole category of findings into false alarms.

Here is my robots.txt:
[PASTE]

For each user agent below, say whether it may fetch [PATHS] and
quote the rule that decides it: Googlebot, Bingbot,
OAI-SearchBot, GPTBot, ChatGPT-User, Claude-SearchBot,
ClaudeBot, Claude-User, PerplexityBot, Google-Extended.

Then say which ones serve search results and which collect
training data, and mark any you are not sure about.

Check: Check the search-versus-training split against each operator's own crawler page, not the model's memory. The usual mistake is a rule meant for training crawlers that also blocks the search crawlers.

Search the web for: [BUYER QUESTION]

Answer it as you normally would. Then list every source you
used with its URL, and for each one say what the page offered
that made it useful: a definition, a number, a comparison
table, a first-hand test, or something else.

Check: Run it in fresh chats with search on, rephrase the question a few times and open every cited URL. Answers change between runs, so one answer is a sample, not a ranking.

Attached are two Search Console page exports: the last three
months, and the same three months a year earlier. Using data
analysis, join them on URL and list pages that lost more than
[X]% of clicks, with clicks, impressions and position for both
periods. Label each page:

- lost position (likely a page problem)
- held position, lost clicks (the results page changed)
- lost impressions (demand fell)

Show the code.

Check: Look at the live results page for the top three before rewriting anything. A page that held position but lost clicks may be losing them to an AI Overview, and the fix there is to be the source it cites.

Fresh chat. You are checking a page before it is published.

1. List every factual claim in the PAGE, TITLE, META
   DESCRIPTION, JSON-LD and ALT TEXT below: numbers, dates,
   prices, names, features, quotes and "studies show" lines.
2. For each claim, quote the sentence in SOURCES that
   supports it, or write NONE. Do not guess.
3. Output a table: claim | supporting quote | keep, soften
   or cut.

PAGE / TITLE / META DESCRIPTION / JSON-LD / ALT TEXT:
[PASTE]

SOURCES:
[PASTE EACH SOURCE WITH ITS URL]

Check: Every NONE row is cut or sourced before publishing. Google calls fact-checking AI-generated content critical and names titles, meta descriptions, structured data and alt text in the same review.