How to Measure AEO Success When AI Answers Keep Changing

The practical framework for measuring visibility, citations, accuracy, traffic, and business impact across ChatGPT, Gemini, Perplexity, Google AI Overviews, and other answer engines.

A question circulating among AEO practitioners captures the problem precisely: run the same prompt in ChatGPT, Perplexity, and Gemini, and each platform may return different sources. Run the same prompt twice, and the citations may change again. If the answer is variable, what does it mean to report that a brand is "visible" in AI search? [1]

The answer is not to abandon measurement. It is to stop treating AEO as a single ranking position.

In short: Measure AEO with a stable, representative prompt set; repeat high-value prompts; track brand presence, competitor visibility, citations, answer accuracy, AI-assisted traffic, and conversions; separate branded from non-branded intent; and report the methodology and uncertainty behind every score.

AEO and GEO are useful terms for monitoring visibility in AI-generated answers. On Google, however, they should be treated as an extension of foundational SEO, not as a replacement for SEO or a separate set of ranking hacks. Google's generative features rely on its existing Search systems, indexed content, and retrieval processes. [9]

AEO success is best measured as a layered system:

1. Exposure: Is the brand present in relevant AI answers?
2. Competitive visibility: Is it present more often, and in better positions, than relevant competitors?
3. Citation quality: Are the brand's pages and trusted third-party sources being used?
4. Brand accuracy: Does the answer describe the brand correctly and consistently?
5. Owned-site behavior: Do AI-referred visitors arrive, engage, and convert?
6. Business impact: Does AI visibility contribute to qualified demand, pipeline, revenue, or retention?

No individual metric answers all six questions. A credible AEO program combines them, reports uncertainty honestly, and connects changes in AI visibility to actions that a marketing team can take.

Why AEO measurement is inherently noisy

Traditional search encourages a simple mental model: one query, one results page, one approximate position. AI search is different. An answer may be produced from several related searches, multiple sources, a model's internal knowledge, the user's conversation history, and the platform's own retrieval and ranking systems.

Google says that AI Overviews and AI Mode may use "query fan-out," meaning the system issues multiple related searches across subtopics and data sources before producing a response. Google also notes that the models and techniques used by AI Overviews and AI Mode can differ, so the responses and links may vary. [4]

That variability creates four common measurement errors.

• Single-run bias: What happens: A marketer records one answer as if it represents the system. Better response: Run repeated observations and report recurrence or ranges.
• Vendor-prompt bias: What happens: A tool's preselected prompts are treated as the complete market. Better response: Build a prompt library from real customer language, funnel stages, locations, and use cases.
• Platform bias: What happens: One model is treated as the entire AI search market. Better response: Track the platforms that matter to the audience and compare them separately.
• Outcome blindness: What happens: Mentions are celebrated without checking accuracy, traffic, leads, or revenue. Better response: Connect visibility data to response quality, analytics, and business outcomes.

This does not make AEO measurement impossible. It changes what a good measurement program looks like. The goal is not to discover one permanent "true" score. The goal is to produce a stable directional signal that is useful for decisions. The 70/20/10 prompt split proposed later in this article is a practical operating heuristic, not an industry standard or a Google recommendation.

Start with a measurement hierarchy

Before choosing a dashboard, define what the organization is trying to learn. AEO metrics should answer progressively harder questions.

1. Are we visible for the questions that matter?

Track the percentage of relevant prompts in which the brand is mentioned or cited. This is a useful presence rate, but it is not a business outcome and should not be reported without its denominator.

For example:

Brand presence rate = prompts with a brand mention ÷ prompts tested

Always report the number of prompts, the platforms tested, the date range, the country or location, and whether the prompt was informational, commercial, local, or branded.

A 60% presence rate across 20 prompts means something different from a 60% rate across 400 prompts. A 60% rate for branded prompts means something different from a 60% rate for non-branded category prompts.

Define the metrics before reporting them

A metric is useful only when two people can calculate it the same way. Define the unit, denominator, evaluator, and reporting window before collecting the baseline.

• Presence rate: Suggested definition: Runs containing a brand mention divided by total runs. Important qualification: A mention may be negative, inaccurate, or irrelevant.
• Citation rate: Suggested definition: Runs containing at least one citation to a specified brand page or domain divided by total runs. Important qualification: A citation indicates retrieval or display, not influence.
• Recommendation inclusion: Suggested definition: Recommendation answers that include the brand divided by relevant recommendation answers. Important qualification: Keep informational, branded, and commercial prompts separate.
• Share of mentions: Suggested definition: Brand mentions divided by all tracked brand mentions in the same answer set. Important qualification: State whether each answer or each mention is the unit of analysis.
• Accuracy rate: Suggested definition: Responses without material or critical factual errors divided by reviewed responses. Important qualification: Publish the truth set and severity rubric used by reviewers.
• AI-assisted conversion rate: Suggested definition: Key events from identified AI-assisted sessions divided by identified AI-assisted sessions. Important qualification: Referrer classification is incomplete, so treat the result as directional.

Do not combine these metrics into a single score unless the weighting is documented and the component metrics remain visible. A small panel can produce large percentage swings. A change from 3 of 5 runs to 4 of 5 runs is not equivalent to a change from 60 of 100 runs to 80 of 100 runs.

2. Are we visible against the right competitors?

Users rarely ask an AI assistant to mention one company in isolation. They ask for recommendations, comparisons, alternatives, best tools, local providers, or solutions for a particular constraint.

Measure brand visibility alongside competitors in the same prompt set. Useful comparative measures include:

• Share of mentions: the proportion of brand mentions across all tracked brands.
• Recommendation inclusion: the percentage of recommendation answers that include the brand.
• Position or prominence: where the brand appears in a list or how prominently it is described.
• Competitive displacement: prompts in which a competitor appears but the brand does not.
• Category coverage: the categories, use cases, locations, and audiences where the brand is or is not present.

Do not interpret a competitor's absence from one answer as a durable win. Competitive visibility is a distribution observed over repeated tests, not a single race result.

3. Are we being cited for the right reasons?

A brand mention without a supporting source may be less useful than a citation to a page that explains the brand's value accurately. Track the sources used in the answer and classify them by role.

• Owned: Examples: Product, service, documentation, pricing, comparison, or location page. What to evaluate: Is the page accurate, current, crawlable, and aligned with the answer?
• Editorial: Examples: Industry publication, review site, analyst article, or news source. What to evaluate: Does the source describe the brand fairly and support the intended positioning?
• Community: Examples: Reddit, forums, social discussion, user-generated reviews. What to evaluate: Is the conversation current, authentic, and consistent with the brand's claims?
• Reference: Examples: Wikipedia, government, standards, or institutional source. What to evaluate: Is the brand represented in authoritative context?
• Competitor or directory: Examples: Marketplace, comparison page, partner listing, or competitor page. What to evaluate: Does the page influence the decision and is the information correct?

The important question is not simply "How many citations did we get?" It is "Which sources are shaping the answer, and which gaps can we address?"

Bing Webmaster Tools' AI Performance report illustrates the distinction. It reports cited pages, cited-page averages, grounding query groupings, and page-level citation activity, while explicitly warning that the data does not measure ranking, authority, importance, or a page's role within an individual answer. The report is aggregated and intended for trend analysis rather than a complete log of every AI answer. [5]

4. Is the AI answer accurate and commercially useful?

Visibility can create risk when the answer is wrong. An AI assistant may confuse a product tier, invent a capability, use an outdated price, attribute a competitor's feature to the brand, or describe a local service as available in the wrong region.

This makes brand accuracy a first-class AEO metric. Create a truth set containing the facts that matter to customers and score each response against it.

• Identity: Does the answer correctly identify the company and category?
• Offering: Are the products or services described accurately?
• Audience: Does the answer identify the right customer or use case?
• Differentiation: Are the stated strengths supported by evidence?
• Availability: Are regions, integrations, plans, and service levels current?
• Commercial details: Are pricing, guarantees, dates, and terms correct?
• Sentiment: Is the overall recommendation positive, neutral, or negative for defensible reasons?
• Citations: Do the cited sources actually support the claims made?

Sentiment is useful, but it should not replace accuracy. A positive answer that contains a material falsehood is a brand-safety problem, not a success. Conversely, a neutral answer may still be valuable if it accurately places the brand in a high-intent comparison.

NIST's generative AI risk guidance recommends context-based measures, continuous monitoring, provenance tracking, human feedback, and documented handling of risks that cannot be measured. Those principles apply directly to AI brand monitoring: define what "correct" means for the business, monitor after changes, and use human review for high-impact errors. [7]

Build a prompt set that represents demand, not a vendor's imagination

A prompt set is a measurement instrument. Its quality determines the usefulness of the result.

Start with actual customer language from sales calls, support tickets, site search, paid-search queries, customer interviews, community discussions, and existing keyword research. Then organize prompts by intent rather than by keyword alone.

• Category discovery: Prompt examples to adapt: "What are the best platforms for [category]?"
• Problem solving: Prompt examples to adapt: "How can a marketing team monitor what AI says about its brand?"
• Comparison: Prompt examples to adapt: "Compare [brand] with [competitor] for a mid-market team."
• Alternative seeking: Prompt examples to adapt: "What are alternatives to [competitor]?"
• Decision stage: Prompt examples to adapt: "Which [service] should a company choose if it needs [constraint]?"
• Proof and risk: Prompt examples to adapt: "Is [brand] reliable, accurate, transparent, or worth the cost?"
• Local intent: Prompt examples to adapt: "Who provides [service] in [city or region]?"
• Branded: Prompt examples to adapt: "What does [brand] do and who is it for?"

For each prompt, record the intent, funnel stage, market, language, platform, model if known, and the expected competitors. Keep a stable core panel for week-over-week comparison. Add a rotating discovery panel to find new questions and emerging competitors.

A practical split is:

• 70% core prompts: stable, high-value questions tracked every reporting period.
• 20% expansion prompts: new variations based on customer language and recent market changes.
• 10% stress-test prompts: ambiguity, misinformation, competitor claims, local edge cases, and high-risk facts.

The exact ratio can change. The principle should not: preserve enough of the panel to detect change while continuing to learn. Do not create pages for every prompt variation. Use prompt research to understand customer needs, then publish only when a page provides substantial value to the underlying audience.

Use repeated runs instead of pretending the answer is deterministic

For high-value prompts, run multiple observations across the reporting period. Report a recurrence rate or interval rather than a binary result.

Mention recurrence = runs containing a brand mention ÷ total runs

For example, if a brand appears in 14 of 20 runs of a commercial comparison prompt, report 70% recurrence across 20 runs, not "the brand ranks." Include the platforms and dates so the result can be interpreted.

Repeat testing is most valuable when it is controlled. Keep the language, location, account state, and prompt wording consistent for the core panel. Record model and platform changes. Do not silently replace prompts that perform poorly; prompt replacement can create artificial improvement.

Measure first-party evidence without overclaiming

AI visibility data tells you what an answer engine displayed. Analytics tells you what happened after a person reached your site. These are related but different layers.

Google Search Central says that traffic from AI Overviews and AI Mode is included in overall Search Console web-search reporting. Google recommends combining Search Console with Analytics and conversion data. [4]

Google Analytics also distinguishes AI Assistant traffic from Google's AI Overviews and AI Mode. The AI Assistant channel can include sources such as ChatGPT, Gemini, DeepSeek, Copilot, and Grok, while Google AI features are classified under Organic Search. [8]

Track at least the following owned-site signals:

• AI-assisted sessions: Why it matters: Shows visits identified as coming from AI assistants or AI-linked surfaces. Main limitation: Referrers can be missing, altered, or classified differently across platforms.
• Engaged sessions: Why it matters: Shows whether visitors found the landing page useful. Main limitation: Engagement is not the same as qualified demand.
• Key events: Why it matters: Connects AI referrals to demo requests, sign-ups, calls, downloads, or purchases. Main limitation: Attribution models distribute credit; they do not prove causality.
• Assisted conversions: Why it matters: Captures AI traffic that contributed before another channel converted the user. Main limitation: Requires consistent event and attribution configuration.
• Landing-page performance: Why it matters: Identifies which pages AI-referred visitors use and whether those pages answer the intent. Main limitation: A page may be valuable even when it receives few direct clicks.
• Direct and self-reported demand: Why it matters: Captures "I heard about you from ChatGPT" or similar signals. Main limitation: Often incomplete and dependent on form quality.

Use a dedicated AI channel or a custom channel group where appropriate. Add self-reported attribution to lead forms, such as "How did you hear about us?" and allow respondents to select AI assistants. Keep the answer optional and treat it as directional evidence.

Do not expect every successful AI interaction to produce a click. Pew Research's analysis of 68,879 Google searches found that traditional result links were clicked in 8% of visits when an AI summary appeared, compared with 15% when it did not. Links within the AI summaries were clicked in only 1% of visits. The study examined Google searches in March 2025 and should not be generalized to every platform, but it demonstrates why clicks alone undercount exposure and influence. [6]

Use a balanced AEO scorecard

A reporting scorecard should combine leading indicators with outcome measures. The following structure is suitable for a monthly marketing review.

• Presence: Recommended metrics: Mention rate, citation rate, recommendation inclusion. Reporting question: Are we appearing for relevant questions?
• Position: Recommended metrics: List position, prominence, share of voice, competitor gap. Reporting question: Are we competitive when we appear?
• Source strength: Recommended metrics: Owned-page citations, third-party citations, source quality, source coverage. Reporting question: What information is influencing the answer?
• Answer quality: Recommended metrics: Accuracy, sentiment, claim support, freshness, severity of errors. Reporting question: Is the answer safe and useful?
• Crawl and content readiness: Recommended metrics: Bot activity, indexed pages, internal links, content coverage. Reporting question: Can relevant systems find and understand the evidence?
• Behavior: Recommended metrics: AI-assisted sessions, engagement, key events, assisted conversions. Reporting question: Do visitors take meaningful actions?
• Business impact: Recommended metrics: Qualified leads, pipeline, revenue, retention, sales feedback. Reporting question: Is AEO contributing to commercial performance?

A dashboard can calculate a composite score, but the component metrics must remain visible. A single score hides trade-offs. Visibility may rise while accuracy falls. Citations may rise while qualified traffic stays flat. AI-assisted sessions may be small while sales teams report a meaningful change in buyer awareness. A citation is evidence of retrieval or display, not proof that the source caused the recommendation, generated a visit, or influenced a purchase. Treat AEO metrics as observational unless you have a controlled test or a clearly defined comparison design.

What not to optimize for

AEO measurement should not become a new form of search-engine-first publishing. Google says there is no need to create special AI files, break content into artificial "chunks," rewrite content solely for AI systems, or create a page for every possible query variation. Google also warns against pursuing inauthentic mentions across the web. [9]

Use prompts to understand customer questions and identify information gaps. Publish content only when it is useful to the audience, original enough to add value, and supported by the organization's actual expertise. This follows Google's people-first guidance: content should provide original analysis, demonstrate relevant expertise, and leave readers satisfied. [11] Correct inaccurate information on legitimate sources, but do not manufacture reviews, forum mentions, citations, or third-party coverage to influence AI answers.

A bot visit is not evidence that a page was cited. A citation is not evidence that a person saw or trusted the page. Treat crawl activity as a diagnostic signal, not as a visibility or performance outcome.

Apply Google's technical baseline

For Google's generative AI features, a page must meet Search technical requirements, be indexed and eligible to appear with a snippet, and the site must be included in Search generative AI features in Search Console where applicable. Eligibility does not guarantee crawling, inclusion, or serving. [9]

The practical baseline is familiar SEO: allow appropriate crawling, use crawlable internal links, make important information available as text, provide a good page experience, keep business details current, and ensure structured data matches the visible page. There is no special schema required for AI Overviews or AI Mode. [4] [9] [10]

For local or service businesses, also maintain accurate business details, a complete Google Business Profile, consistent service-area information, legitimate reviews, and appropriate local or service structured data where it reflects visible content. Use Merchant Center when product information is relevant. Do not create fake locations or duplicate location pages solely to capture query variants. [9]

A practical measurement record

A repeatable measurement process needs a repeatable record. For each observation, capture at least the following fields.

• Prompt ID: COMP-014
• Prompt text: Best AI visibility tools for B2B marketing teams
• Intent: Commercial comparison
• Market: United States
• Platform and model: ChatGPT; model recorded if available
• Run date: 2026-09-13
• Brand mentioned: Yes or no
• Brand position: 2, if the answer uses an ordered list
• Citation present: Yes or no
• Citation URL: The cited page, if available
• Sentiment: Positive, neutral, or negative
• Accuracy severity: None, minor, material, or critical
• Competitors mentioned: Names of competing brands
• Reviewer notes: Claim-support and context notes

This record makes results auditable. It also prevents a dashboard from hiding changes in the prompt set, platform, location, model, or evaluator.

Worked example: interpreting a noisy result

Suppose a team tests 25 non-branded comparison prompts across three platforms, with two runs for each high-value prompt. The brand appears in 24 of 50 runs, giving a 48% presence rate. It is cited in 15 runs, giving a 30% citation rate. Eight of those citations point to the company's own pages, while seven point to third-party sources. Three responses contain material errors about pricing or availability.

The correct conclusion is not "the brand ranks at 48%." A more useful conclusion is:

The brand appeared in nearly half of the observed runs, but only one-third included a citation. Owned-source coverage is incomplete, and three material accuracy errors require remediation before visibility is treated as a success.

The next actions would be to improve the pages that answer the comparison intent, correct the conflicting commercial facts, and retest the same prompt panel after an appropriate refresh period. This is an illustrative example, not a reported Anny case study.

A practical 30-day AEO measurement program

Week 1: Define the market and the truth set

Select the business categories, customer segments, locations, competitors, and decisions that matter. Build the core prompt panel from real customer language. Create the brand truth set for identity, offering, audience, differentiation, availability, commercial details, and claims.

Document the measurement conditions. Record platforms, dates, locations, languages, account states, model names when available, and sampling rules.

Week 2: Establish the baseline

Run the core panel across the selected platforms and repeat high-value prompts. Save the full answers, citations, screenshots or exports where permitted, and evaluator notes. Tag each result by prompt intent and funnel stage.

Set up or verify Search Console, Analytics, conversion events, AI referral classification, server or CDN logs, and lead-source questions. Use Anny to consolidate response monitoring, visibility, sentiment, competitive comparisons, sources, models, and AI crawler activity in one operating view. [2]

Week 3: Diagnose gaps

Separate four types of gap:

1. Presence gap: the brand does not appear for a relevant question.
2. Source gap: the answer relies on weak, outdated, or incomplete sources.
3. Accuracy gap: the brand appears but is described incorrectly.
4. Conversion gap: the brand appears and receives visits, but the landing experience does not convert the intent.

Each gap requires a different intervention. A presence gap may require clearer category and use-case content. A source gap may require better owned pages, expert contributions, reviews, partnerships, or public documentation. An accuracy gap requires correcting inconsistent facts on important, legitimate sources and making the authoritative version clear on the company's own site. A conversion gap requires improving the destination page and measurement path.

Week 4: Implement, retest, and report

Prioritize changes by commercial value and error severity. Update pages so important claims are explicit, current, text-accessible, and supported by evidence. Improve internal linking and ensure structured data matches visible content. Google specifically advises maintaining foundational SEO practices, crawlability, internal links, textual content, page experience, and consistency between structured data and visible content. [4] [10]

Retest the affected prompt clusters. Do not expect every content change to produce an immediate shift. AI systems recrawl and refresh on different schedules, and a citation change may reflect a model or query change rather than the content edit itself.

Report the baseline, changes made, observed movement, confidence level, and next action. Use language such as "citation recurrence increased across the core commercial panel" rather than "we moved to position one."

What Anny can help a marketing team do

Anny's value is strongest when it is used as part of this measurement system rather than as a replacement for judgment.

AI visibility monitoring helps teams observe how often brands appear across major AI platforms. Source analysis helps reveal which domains and pages are shaping answers. Competitive comparison shows where competitors are present and where the brand is missing. Response and sentiment monitoring helps identify inaccurate, negative, or inconsistent descriptions. Crawler and query insights provide clues about what content systems are accessing and which topics are associated with citations. [2]

For teams that need implementation support, Anny's services extend beyond reporting. Custom strategy can translate the baseline into a prioritized plan. Fully managed execution can support content optimization and external visibility work. Ongoing performance audits can identify changes as platforms evolve. Training and workshops can help in-house teams build a repeatable GEO process. [3]

The most useful engagement is therefore not "make our score higher." It is:

Help us become more accurately visible for the questions that influence our customers, then prove whether that visibility contributes to demand.

Common AEO measurement mistakes to avoid

Mistake 1: Reporting a score without a denominator

"Visibility increased by 12%" is incomplete. State how many prompts were tested, which platforms were included, and how the metric was calculated.

Mistake 2: Treating one answer as ground truth

A single response is an observation. It is not a ranking position, market share, or stable representation of all users.

Mistake 3: Optimizing for mentions alone

A mention can be negative, inaccurate, irrelevant, or buried in a low-intent answer. Pair presence with position, sentiment, accuracy, citations, and outcomes.

Mistake 4: Mixing branded and non-branded prompts

Branded prompts measure recognition and entity understanding. Non-branded prompts measure category discoverability. Keep them separate.

Mistake 5: Confusing citations with influence

A cited page may appear in an answer without driving a click or affecting the recommendation. Citation data is valuable evidence, but it is not proof of causality.

Mistake 6: Treating platform and tool data as complete market data

Platform reports and third-party tools observe only the surfaces, prompts, samples, and time periods they can access. Search Console, analytics platforms, crawler logs, and synthetic prompt tools answer different questions. Do not combine them without documenting the normalization method and limitations.

Mistake 7: Ignoring source consistency

AI systems synthesize information from across the web. Conflicting company names, descriptions, pricing, locations, reviews, or product claims can reduce answer quality. Maintain a consistent, verifiable entity profile across important sources.

Mistake 8: Changing the prompt set whenever the results are inconvenient

Prompt churn destroys comparability. Protect the core panel and document every addition, removal, or wording change.

Measurement limitations to disclose

AEO data is observational and changes as platforms, models, retrieval systems, prompts, locations, and source pages change. Synthetic prompts do not represent every real user conversation. Referrer data can be missing or classified inconsistently. Search Console and platform dashboards may aggregate or sample activity. Sentiment and accuracy judgments can vary between evaluators.

For these reasons, report trends with their sample size, platform scope, time window, prompt methodology, and known limitations. Use AEO data to prioritize decisions and investigate changes. Do not present it as a complete census of everything an AI system says or as proof that a particular optimization caused revenue.

The standard to aim for

A mature AEO program does not claim perfect visibility measurement. It makes uncertainty explicit and still produces better decisions.

The standard is a repeatable panel of real buyer questions, tested across relevant platforms and markets. It is a record of full responses and citations, not just a score. It measures whether the answer is accurate and useful, not merely positive. It connects AI observations to first-party behavior, qualified demand, and revenue without overstating attribution. It monitors change over time and preserves enough methodological consistency to distinguish a real signal from model noise.

AI answer variability is not a reason to stop measuring. It is a reason to measure more intelligently.

If you want to understand what AI says about your brand, which sources it trusts, and where your competitors are being recommended instead, start with an AI visibility audit from Anny.

Editorial note: This guide was published by Anny, an AI search analytics and AEO services provider. The framework distinguishes platform observations from business outcomes and does not guarantee rankings, citations, traffic, or conversions.

References

• [1] How are you actually measuring AEO success? Every tool I've tested gives me different answers for the same prompt: https://www.reddit.com/r/aeo/comments/1t7rlna/how_are_you_actually_measuring_aeo_success_every/
• [2] Anny: AI search analytics for marketing teams: https://anny.dodoxhq.com/
• [3] Anny services: Monitor and boost your brand's visibility on AI search: https://anny.dodoxhq.com/services
• [4] AI features and your website, Google Search Central: https://developers.google.com/search/docs/appearance/ai-features
• [5] AI Performance in Bing Webmaster Tools: https://www.bing.com/webmasters/help/ai-performance-9f8e7d6c
• [6] Google users are less likely to click on links when an AI summary appears in the results, Pew Research Center: https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
• [7] Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
• [8] Default channel group and traffic-source dimensions, Google Analytics Help: https://support.google.com/analytics/answer/9756891
• [9] Optimizing your website for generative AI features on Google Search, Google Search Central: https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
• [10] Google Search Essentials: https://developers.google.com/search/docs/essentials
• [11] Creating helpful, reliable, people-first content, Google Search Central: https://developers.google.com/search/docs/fundamentals/creating-helpful-content