How to Track AI Visibility for SEO

AI visibility tracking should show where a brand appears, which sources are cited, how competitors are presented, and whether the answers are accurate. The useful part is not collecting random screenshots. It is checking a stable set of commercially relevant prompts, connecting the results to pages and tasks, and reviewing the same signals over time.

TL;DR

• Build a stable prompt set around customer problems, categories, comparisons, products, and branded questions.
• Track ChatGPT, Google AI Overviews, Gemini, Perplexity, Copilot, or the platforms that matter to the audience.
• Record brand mentions, citations, cited URLs, competitors, answer accuracy, prompt intent, platform, location, and date.
• Separate broad prompt tracking from custom prompts tied to important products or services.
• Use Ahrefs Brand Radar, Semrush AI visibility features, or specialist tools when manual checking becomes too slow.
• Use Screaming Frog and technical checks to confirm that important pages, structured data, sources, and bot access remain available.
• Keep AI visibility tasks inside the normal SEO workflow, then connect the implemented page changes to later prompt reviews.
• Do not claim that one content edit caused an AI-answer change. Track repeated patterns and supporting evidence.

Start with the business questions that matter

Do not begin with hundreds of prompts because an AI visibility tool allows them.

Start with questions that could influence:

• Product discovery
• Vendor shortlists
• Category awareness
• Brand comparison
• Purchase decisions
• Problem research
• Trust and credibility
• Customer support
• Reputation
• Existing customer understanding

For an SEO software company, useful questions may include:

What are the best tools for tracking SEO work?
How should an SEO agency document completed work?
Which tools monitor important SEO page changes?
What is the best alternative to managing SEO tasks in spreadsheets?
How can an agency prove what SEO work was completed?

For an e-commerce brand:

What are the best [product category] for [use case]?
Which [product] is best for [customer type]?
What are the alternatives to [competitor]?
Is [brand] good for [specific problem]?
How does [brand] compare with [competitor]?

The prompt set should reflect actual buying and research behavior, not only keywords with search volume.

Divide prompts by intent and topic

Use prompt groups so the results can be reviewed as a pattern.

Keep the group names stable. Changing the category structure every month makes trend analysis difficult.

A practical starting set:

• 5 to 10 category prompts
• 5 to 10 problem-based prompts
• 5 comparison or alternative prompts
• 5 branded and reputation prompts
• 5 educational prompts where citation visibility matters

That produces a manageable set of 25 to 40 prompts.

Keep the exact prompt wording stable

Small wording changes can produce different answers.

These prompts are related but not identical:

Best SEO tools for agencies
Best tools for managing SEO agency work
What software should an SEO agency use?
Which tools help agencies prove completed SEO work?

Track the exact wording used in each check.

For each prompt, store:

• Prompt ID
• Prompt group
• Exact prompt
• Business priority
• Target market
• Language
• Location
• Device or environment where relevant
• Platform
• Check date
• Reviewer

Do not rewrite the prompt after a poor result and treat the new answer as an improvement. Add the new wording as a separate prompt or record the change in the tracking history.

Choose the platforms the audience actually uses

A practical cross-platform set may include:

• ChatGPT Search
• Google AI Overviews
• Gemini
• Perplexity
• Microsoft Copilot

Some teams may also review:

• Google AI Mode
• Claude when the use case involves web research
• Industry-specific answer engines
• E-commerce assistants
• Marketplace or travel recommendation systems

Do not assume all platforms will produce the same result.

They may differ because of:

• Search and retrieval systems
• Available sources
• Location
• Personalization
• Model version
• Query rewriting
• Freshness
• Platform-specific ranking and citation behavior
• Whether live web search is used

ChatGPT Search may include inline citations and a source panel when web search is used. Google AI Overviews can also surface supporting links in Search. Track both the generated answer and the cited sources rather than recording only whether the brand name appeared.

Record more than a yes-or-no mention

A useful tracking row contains enough detail to explain the result later.

Useful controlled values reduce inconsistent reporting.

Brand mentioned

• Yes
• No

Brand cited

• Yes
• No
• Mentioned without citation
• Cited through a third-party source

Description accuracy

• Accurate
• Mostly accurate
• Partially accurate
• Misleading
• Incorrect
• Not enough detail

Recommendation strength

• Primary recommendation
• Included in shortlist
• Mentioned as an alternative
• Mentioned without recommendation
• Not mentioned

Sentiment

• Positive
• Neutral
• Negative
• Mixed

Store the answer evidence

AI answers can change. A later reviewer should be able to see what was measured.

Store one or more of:

• Screenshot
• Export from the visibility tool
• Saved response text
• Source URLs
• Platform conversation link where appropriate
• Date and time
• Market and language
• Model or product label if shown

Do not store sensitive customer information in prompts or screenshots.

A simple file naming rule:

2026-07-12_chatgpt_category_best-seo-work-tracking-tools.png

For agencies:

client_project_date_platform_prompt-id

Example:

acme_ai-visibility_2026-07-12_perplexity_CAT-004.png

Use a spreadsheet for the first prompt set

A spreadsheet is enough while the prompt set is small and the checks are manual.

Recommended tabs:

Prompt Library

Checks

Work

The SEO tracking spreadsheet template already includes an AI Visibility tab that can be adapted for this workflow.

Calculate useful AI visibility metrics

Do not rely on one combined score without understanding the inputs.

Mention rate

Prompts where the brand appeared ÷ Prompts checked

Example:

18 mentions ÷ 40 prompts = 45% mention rate

Citation rate

Prompts where the brand or site was cited ÷ Prompts checked

Example:

8 citations ÷ 40 prompts = 20% citation rate

Official-domain citation rate

Prompts citing the official site ÷ Prompts checked

This separates direct citations from answers that mention the brand through third-party sources.

Shortlist rate

Prompts where the brand appeared as a recommendation or shortlist option ÷ Prompts checked

Accuracy rate

Accurate or mostly accurate brand descriptions ÷ Prompts where the brand appeared

Competitor share of mentions

Competitor mentions ÷ Total tracked brand mentions

Use this carefully because an answer may mention several brands.

Prompt coverage

Prompt groups with at least one brand mention ÷ Total prompt groups

A brand may have a strong mention rate in branded prompts but remain absent from category and problem-based prompts. Group-level reporting makes that visible.

Report metrics by prompt group and platform

An overall rate can hide the real issue.

Example:

This tells the team:

• The brand is recognized when named.
• It is rarely included in broader category questions.
• Educational content earns some citations even when the product is not mentioned.
• Comparison descriptions need correction.

Also compare platforms:

Do not present the percentages without the prompt count.

“50% visibility” may mean 2 of 4 prompts or 500 of 1,000 prompts.

Review the sources behind the answer

A brand can appear because the platform cites:

• Official product pages
• Documentation
• Articles
• Research
• Review platforms
• Reddit
• YouTube
• News sites
• Competitor comparison pages
• Directories
• Marketplace listings

Track:

• Source domain
• Source URL
• Official or third party
• Page type
• Whether the source is accurate
• Whether it is current
• Which prompts cite it
• Which competitors it mentions
• Whether the team controls the source

A source-level table:

Do not treat every third-party citation as a link-building target. First understand why the page is useful to the answer.

Track competitors as entities, not one text field

A single “competitors mentioned” cell becomes difficult to analyze.

For serious tracking, store one competitor per row in a related table:

This allows the team to calculate:

• Competitor mention frequency
• Share of shortlist appearances
• Citation frequency
• Topics each competitor owns
• Platforms where a competitor is strongest
• Sources that repeatedly support the competitor

Do not assume the competitor with the highest mention count has the strongest commercial position. Review recommendation strength and prompt intent.

Use Ahrefs, Semrush, or a specialist tool when manual checks stop scaling

Manual tracking works for a small stable prompt set. A visibility platform becomes useful when the team needs:

• More prompts
• More brands
• More markets
• More languages
• More frequent checks
• Historical trends
• Competitor comparisons
• Citation analysis
• Exports
• Team dashboards
• Alerts

Ahrefs Brand Radar supports brand and competitor visibility analysis across AI answers and offers custom prompt tracking. Semrush also provides AI visibility functions as part of its broader search and marketing toolset.

Specialist platforms may focus more deeply on:

• Prompt tracking
• AI share of voice
• Citation discovery
• Answer sentiment
• Competitor benchmarking
• Market and language coverage
• Team reporting

Before buying, test:

1. Which platforms are supported?
2. Are the prompts synthetic, search-backed, custom, or a mixture?
3. Can the exact prompt be viewed?
4. Can markets and languages be controlled?
5. Can response evidence be exported?
6. Are citations and source URLs included?
7. How is share of voice calculated?
8. Can competitors be edited?
9. How often are prompts rechecked?
10. Can data be exported by prompt and date?
11. Does the tool preserve answer history?
12. Can findings connect to the team’s normal SEO workflow?

Use Screaming Frog for the technical and page-level checks

Screaming Frog cannot tell you whether a brand was recommended in ChatGPT or cited in Perplexity. It can confirm whether the pages intended to support that visibility still contain the required signals.

Useful checks include:

• Status codes
• Indexability
• Canonicals
• Titles and headings
• Author information
• Structured data
• sameAs references
• Internal links
• External citations
• Publication and update dates
• Product descriptions
• Comparison content
• Content blocks
• Robots directives
• AI bot access where it can be tested from the site configuration

Use custom extraction for fields such as:

• Author name
• Expert credentials
• Lead statistic
• Source list
• Product summary
• Competitor table
• Organization description

The process in How to Compare Screaming Frog Crawls can confirm whether those page elements changed between releases.

The LLM Readiness Checklist can also help review the technical, content, source, and entity signals worth checking before treating an AI visibility problem as a prompt-only issue.

Check access for search and AI crawlers

A page cannot be retrieved through a search-based AI experience if relevant crawlers and search systems cannot access it.

For ChatGPT Search, OpenAI states that sites should allow OAI-Searchbot and the published IP ranges used for search access. Review:

• Robots.txt
• CDN rules
• Firewall rules
• Bot-management systems
• Authentication
• JavaScript rendering
• Server responses
• Rate limits
• Geolocation restrictions
• Accidental blocking by security services

Do not assume that adding llms.txt fixes blocked crawling, poor indexability, weak source quality, or unclear content.

A practical technical review:

The free Indexability Checker can handle a one-time page review. Important pages should then be monitored according to risk.

Connect prompt findings to pages and tasks

A useful AI visibility workflow looks like this:

Prompt check
→ Visibility or accuracy gap
→ Related page or source identified
→ Task assigned
→ Change implemented
→ Page verified
→ Prompt set rechecked later
→ Result recorded

Example:

Do not create vague tasks such as:

Improve GEO

Use specific work:

Update /platform/seo-work-tracking with:
- clearer category description
- agency workflow example
- comparison with general PM tools
- evidence of URL monitoring
- links to technical documentation

Keep AI visibility work inside the normal SEO history

AI-related work can include:

• Updating product descriptions
• Clarifying company and entity information
• Improving author pages
• Adding original research
• Adding or correcting sources
• Updating comparison content
• Creating documentation
• Improving structured data
• Fixing bot access
• Strengthening internal links
• Correcting inaccurate public information
• Updating old third-party profiles
• Creating a new page for a missing customer problem

Record the same fields used for other SEO work:

• URL
• Date
• Owner
• Change
• Reason
• Related prompt group
• Verification
• Review date
• Outcome
• Evidence

SEO Logbook can keep that work beside traditional SEO changes, monitoring detections, tasks, and later impact notes.

A URL history may show:

This is more useful than a visibility chart with no record of what the team changed.

Do not use one recheck as proof of impact

AI answers are variable.

A changed answer may result from:

• Different retrieval
• Freshness
• Model updates
• Location
• Personalization
• Query rewriting
• New third-party content
• Competitor activity
• Search-index changes
• Platform experiments
• Random response variation

Use repeated checks and cautious outcome labels:

• Positive signal
• Negative signal
• No clear change
• Mixed across platforms
• Accuracy improved
• Citation gained
• Citation lost
• Needs more time
• Confounded by external changes

Track referral traffic separately

Some AI experiences send referral traffic. Others may influence awareness without producing a click.

In analytics, review:

• Referral source
• Landing page
• Sessions
• Engaged sessions
• Conversions
• Assisted conversions
• New users
• Revenue where appropriate

Keep referral reporting separate from prompt visibility.

A platform can show:

• Higher mention rate
• Higher citation rate
• No measurable traffic

Or:

• Low prompt visibility in the tracked set
• Meaningful referral traffic from a cited guide

Both observations matter.

Do not use referral traffic as the only AI visibility metric because many answers do not generate a visit.

Analyze a manual CSV with Python

Use a CSV with columns such as:

check_date
platform
prompt_id
prompt_group
brand_mentioned
brand_cited
official_domain_cited
accuracy
recommendation_strength
competitors
citation_url
related_url
related_work_id

Example Python analysis:

from __future__ import annotations

from pathlib import Path

import pandas as pd

INPUT_FILE = Path("ai-visibility-checks.csv")
OUTPUT_FILE = Path("ai-visibility-summary.csv")

YES_VALUES = {"yes", "true", "1", "y"}


def to_bool(series: pd.Series) -> pd.Series:
    return (
        series.fillna("")
        .astype(str)
        .str.strip()
        .str.lower()
        .isin(YES_VALUES)
    )


def main() -> None:
    checks = pd.read_csv(INPUT_FILE)

    required = {
        "platform",
        "prompt_id",
        "prompt_group",
        "brand_mentioned",
        "brand_cited",
        "official_domain_cited",
    }

    missing = required - set(checks.columns)

    if missing:
        raise ValueError(
            f"Missing required columns: {sorted(missing)}"
        )

    checks["brand_mentioned_bool"] = to_bool(
        checks["brand_mentioned"]
    )
    checks["brand_cited_bool"] = to_bool(
        checks["brand_cited"]
    )
    checks["official_domain_cited_bool"] = to_bool(
        checks["official_domain_cited"]
    )

    summary = (
        checks.groupby(
            ["platform", "prompt_group"],
            dropna=False,
        )
        .agg(
            prompts_checked=("prompt_id", "nunique"),
            checks=("prompt_id", "size"),
            mentions=("brand_mentioned_bool", "sum"),
            citations=("brand_cited_bool", "sum"),
            official_citations=(
                "official_domain_cited_bool",
                "sum",
            ),
        )
        .reset_index()
    )

    summary["mention_rate"] = (
        summary["mentions"] / summary["checks"]
    )
    summary["citation_rate"] = (
        summary["citations"] / summary["checks"]
    )
    summary["official_citation_rate"] = (
        summary["official_citations"] / summary["checks"]
    )

    rate_columns = [
        "mention_rate",
        "citation_rate",
        "official_citation_rate",
    ]

    summary[rate_columns] = (
        summary[rate_columns] * 100
    ).round(1)

    summary = summary.sort_values(
        ["platform", "prompt_group"],
        kind="stable",
    )

    summary.to_csv(OUTPUT_FILE, index=False)

    print(summary.to_string(index=False))
    print(f"\nSaved: {OUTPUT_FILE.resolve()}")


if __name__ == "__main__":
    main()

Install pandas:

python -m pip install pandas

Run:

python analyze_ai_visibility.py

The output provides prompt counts and rates by platform and prompt group.

Do not combine results from different months into one rate unless that is the intended reporting period.

Expand competitor analysis from the same CSV

If the competitors column uses semicolon-separated names:

Competitor A; Competitor B; Competitor C

Use:

competitors = (
    checks.assign(
        competitor=checks["competitors"]
        .fillna("")
        .str.split(";")
    )
    .explode("competitor")
)

competitors["competitor"] = (
    competitors["competitor"]
    .astype(str)
    .str.strip()
)

competitors = competitors[
    competitors["competitor"] != ""
]

competitor_summary = (
    competitors.groupby(
        ["platform", "prompt_group", "competitor"]
    )
    .size()
    .reset_index(name="mentions")
    .sort_values(
        ["platform", "prompt_group", "mentions"],
        ascending=[True, True, False],
    )
)

competitor_summary.to_csv(
    "ai-competitor-mentions.csv",
    index=False,
)

For a stronger system, use a separate related table instead of a semicolon-separated cell.

Build a monthly review that stays practical

Review five areas.

1. Prompt coverage

• Which prompt groups include the brand?
• Which groups remain absent?
• Did category coverage change?
• Are branded answers accurate?

2. Citations

• Which official pages are cited?
• Which third-party pages shape the answer?
• Were citations gained or lost?
• Are cited pages still current and available?

3. Competitors

• Which competitors appear most often?
• Which topics do they own?
• Which sources support them?
• Are they recommended or only mentioned?

4. Work completed

• Which pages were updated?
• Which technical access issues were fixed?
• Which sources or author pages were improved?
• Which AI visibility tasks remain open?

5. Outcomes and next actions

• Which changes show a positive signal?
• Which results remain mixed?
• Which descriptions are still inaccurate?
• Which prompt groups need more evidence or a better page?
• Which checks should continue next month?

A practical report table:

Avoid the common tracking mistakes

Random prompts every month

A changing prompt set makes trend comparisons unreliable.

Tracking only branded questions

Branded prompts test recognition and accuracy. Category and problem prompts test discovery.

Using mention rate without citations

A brand can be named through an inaccurate third-party source.

Using citations without answer context

A page may be cited while a competitor receives the recommendation.

Combining all platforms

Platform-level differences can disappear inside one overall score.

Ignoring location and language

The answer can change by market.

Reporting one answer as a stable ranking

AI answers are not fixed SERP positions.

Creating tasks with no related page

The team cannot act on “improve AI visibility” without a specific asset or source problem.

Tracking page work without verification

The planned content may never reach production or may later be overwritten.

Treating llms.txt as a complete strategy

Technical access does not replace useful content, clear entity information, credible sources, and discoverable pages.

Buying a tool before defining the prompt set

The software cannot decide which customer questions matter to the business.

Use this implementation checklist

Start with a small set that the team can check consistently. A controlled 30-prompt review with clear ownership is more useful than a 5,000-prompt dashboard that nobody connects to actual SEO work.

FAQs

How many prompts should a team track?

Start with 25 to 40 stable prompts across category, problem, comparison, branded, and educational groups. Add more when the team can maintain the checks and act on the findings.

How often should AI visibility be checked?

Monthly is a practical starting point for a stable prompt set. High-priority launches, reputation issues, or active tests may justify weekly checks. Avoid reacting to daily variation without a clear reason.

Which metrics should be reported?

Use prompt count, mention rate, citation rate, official-domain citation rate, shortlist rate, accuracy, competitor visibility, cited sources, AI referral traffic, and related work completed.

Can Google Search Console track AI visibility?

Use GSC for Google Search clicks, impressions, queries, pages, and indexing evidence. Keep prompt-level mentions, answer accuracy, citations, and cross-platform visibility in a separate tracking system.

Does Screaming Frog track ChatGPT or AI Overview mentions?

No. Screaming Frog can verify the technical and page-level signals that support important content, such as indexability, headings, structured data, sources, and author details. Use prompt tracking or an AI visibility platform for answer-level measurement.

How should agencies connect AI visibility to client work?

Store the client, prompt group, platform, related URL, identified gap, assigned task, implemented change, verification, recheck date, and later observation. Report trends and evidence without claiming unsupported attribution.