When Google AI Mode produces a detailed answer followed by 10, 15 or even 20 source links, the response can appear extremely well researched.

The natural assumption is that the citations prove the claims placed beside them.

But do they?

We conducted a simple experiment to find out.

Rather than merely counting the number of citations or looking at which websites appeared, we took a series of AI Mode answers and checked the claims against the cited pages one by one.

The results were more complicated than a simple verdict of “accurate” or “inaccurate.”

AI Mode frequently found relevant sources and produced practical, coherent advice. However, the citations did not always support the exact wording, certainty, numerical precision or final recommendation presented in the answer.

In several places, AI Mode appeared to:

Combine different recommendations into one new rule.
Strengthen cautious source wording.
Move advice from one context into another.
Present disagreement as consensus.
Use genuine scientific research to support conclusions that the research had not directly tested.

For hobby website builders, this matters enormously.

It helps explain both what AI Mode can already do very well and where a carefully researched, evidence-led hobby website can still add something genuinely useful.

What we wanted to test

The experiment was built around one question:

When AI Mode makes a claim and places citations nearby, do those cited pages directly support that claim?

This is different from asking whether the overall answer sounds sensible.

An answer can be practical and broadly correct while still presenting its citations imprecisely.

For example, imagine AI Mode says:

Memorising the first 60 seconds of a presentation reduces anxiety.

It might cite pages that recommend preparing an opening, pages that discuss public-speaking anxiety and pages that warn against memorising an entire speech.

Taken together, those sources may point towards a reasonable compromise.

But that does not necessarily mean any of them established that:

Exactly 60 seconds should be memorised.
Memorisation directly reduces anxiety.
The recommendation applies to every nervous beginner.
This method is superior to all alternatives.

Our experiment therefore examined the connection between the specific claim and the specific evidence.

Why we chose public speaking

We deliberately chose a subject where credible coaches hold different opinions.

Some public-speaking specialists recommend memorising at least part of the opening. Others warn that word-for-word memorisation can make a speaker sound unnatural or cause problems if the memorised sequence is interrupted.

The disagreement makes public speaking more revealing than a straightforward factual query.

A question such as how many centimetres are in a metre has one stable answer.

A question about how a nervous beginner should prepare a presentation involves:

Individual differences.
Practical experience.
Coaching judgment.
Psychological claims.
Different speaking situations.
Different definitions of memorisation.
A genuine difference of professional opinion.

This created an opportunity to see whether AI Mode would preserve that complexity or compress it into a single clean recommendation.

The experiment

We asked three related questions.

Each new query was designed to place more pressure on the original answer.

Query one

Should a nervous beginner memorise the first 60 seconds of a presentation, or use an outline instead?

Query two

For a nervous first-time speaker, is it better to memorise the opening of a presentation or rely entirely on brief notes?

Query three

Some public-speaking coaches recommend memorising the opening of a presentation, while others warn against memorisation. What are the strongest arguments on both sides, and what does the available evidence actually establish?

We copied each AI Mode response in full, including its numbered citations.

We then:

Divided the answer into individual claims.
Opened the cited pages.
Checked what each source actually said.
Compared the source wording with AI Mode’s wording.
Looked for disagreement between the cited sources.
Separated direct source support from AI Mode’s own synthesis.

How we classified the claims

We used several simple categories.

Direct support

The cited page clearly supported the complete claim in substantially the same context.

Partial support

The source supported part of the claim but not its full wording, precision or certainty.

Reasonable inference

The source supplied relevant information, but AI Mode had taken an additional reasoning step.

Combined support

The claim appeared to have been created by combining several sources.

Context drift

The source supported the advice in one setting, but AI Mode applied it in another.

Unsupported

We could not find the claim in the cited material.

Contradicted

The cited source said something materially different from the AI Mode statement.

These distinctions matter because “related to the subject” is not the same as “proves the claim.”

Run One: The 60-second compromise

The first response opened with a clear recommendation:

“A nervous beginner should memorise the first 60 seconds of a presentation and use a bulleted outline for the remaining duration.”

It then explained that memorising the opening would:

Bypass the initial adrenaline response.
Build early momentum.
Improve eye contact.
Create a strong first impression.

It warned that memorising the entire presentation could cause the speaker to go completely blank after forgetting a word.

The final practical recommendation was:

Write the first 60 seconds word for word.
Use bold keywords for the body.
Write the final line word for word.

The answer was clear, usable and confident.

The problem was that the cited sources did not establish this exact formula.

The sources recommended different opening lengths

One cited Universal Class page recommended having the first 20 seconds completely memorised, largely to support immediate eye contact. (Universal Class: Ways to Rehearse Your Speech)

Another source, 98th Percentile, recommended a hybrid approach: memorise key points or the introduction and conclusion while using concise notes for the body. (98th Percentile: Memorizing Your Speech)

Mentimeter advised speakers to “master the first 60 seconds,” but its method was to start with a strong hook and become comfortable with the material. The same page explicitly advised:

“Practice, don’t memorize.”

It did not recommend writing and memorising the opening word for word. (Mentimeter: How to Do a Presentation)

AI Mode had therefore combined several related recommendations:

Memorise 20 seconds.
Prepare the introduction carefully.
Use notes for the body.
Master the first 60 seconds.
Avoid memorising the entire speech.

From those pieces, it created one precise rule:

Memorise exactly 60 seconds, then switch to an outline.

That rule may be sensible.

But it was AI Mode’s synthesis, not a single recommendation directly established by the cited sources.

One citation contradicted the nearby instruction

The note-structure section cited only the Mentimeter page.

AI Mode said:

“Write out your hook, introduction, and core message word-for-word.”

But Mentimeter said:

“Practice, don’t memorize.”

The page supported giving particular attention to the first minute, beginning with a hook, maintaining eye contact and finishing with a clear call to action.

It did not support memorising the first minute word for word.

This was not just weak support. It was a meaningful difference between the advice in the answer and the advice in the cited page.

The citations were attached to sections rather than claims

AI Mode placed citations [1–5] after a group of four claims and citations [6–10] after another group of three.

This created an additional problem.

The reader could not tell:

Which source supported which bullet.
Whether all five sources supported every bullet.
Whether some sources disagreed.
Which statements were direct source findings.
Which statements were AI Mode’s own conclusions.

The citation group gave a broad impression of support without providing sentence-level traceability.

Run Two: The answer becomes more confident

We then rephrased the query:

For a nervous first-time speaker, is it better to memorise the opening of a presentation or rely entirely on brief notes?

AI Mode responded:

“Memorising the first 60 to 90 seconds of your presentation is vastly superior to relying entirely on notes.”

The broad recommendation remained the same, but two things changed.

First, the prescribed opening grew from 60 seconds to 60–90 seconds.

Second, the recommendation became much stronger:

“Vastly superior.”

Most of the cited domains also changed. Only Winning Presentations, Universal Class and 98th Percentile appeared in both source sets.

This gave us our first major pattern:

The recommendation remained stable while the supporting citation set changed substantially.

A source became the apparent anchor

The Effective Presentations article cited first in the second response discusses problems with full-script memorisation, describes the first 90 seconds as a particularly difficult period and recommends memorising the opening and closing while mastering the concepts and structure of the body. (Effective Presentations: Why Memorizing Your Speech Backfires)

This was a much closer match to AI Mode’s central recommendation.

However, the page still did not establish that the approach was:

“Vastly superior”

for every nervous first-time speaker.

That phrase was AI Mode’s judgment.

New precision appeared without clear evidence

The second answer added a detailed execution plan:

Write exactly three to five sentences.
Keep each sentence under ten words.
Develop a verbal bridge into the first main point.
Use thick card stock.
Have three to four bullet points per slide.
Use large, bold text.
Highlight keywords in colour.

The answer made these instructions look like one coherent expert system.

But the supporting links came from several different settings, including:

Public-speaking coaching.
Slide-design advice.
A Reddit discussion.
A general university note-taking article.

Some individual tips may have been sensible.

The citation problem was that they had been gathered from different contexts and presented as one unified method for creating speaker notes.

For example, advice about limiting the number of bullet points on a slide is not necessarily evidence for the correct number of bullet points on a private cue card.

Advice about colour-coding general study notes is not automatically evidence that colour-coded presentation cards improve public-speaking performance.

This is what we called context drift.

The original source may genuinely contain the advice, but AI Mode moves it into a new setting without clearly marking that reasoning step.

The important question was not whether the tips were useful

Some of the practical recommendations may work very well.

That was not what we were testing.

We were testing whether the citations demonstrated that the recommendations had:

Been tested in that context.
Been recommended by the cited expert for that purpose.
Been shown to be better than alternatives.
Been established as general rules.

The answer often moved from:

“Here is a related idea found in a source”

to:

“Here is the correct method to follow.”

Run Three: Asking directly for the evidence

The third query explicitly identified the disagreement and asked what the evidence established.

This time, AI Mode acknowledged the debate:

“Memorising the exact opening lines of a presentation is highly debated among public-speaking coaches.”

That was progress.

It then presented arguments for and against memorisation before offering a section titled:

“What the Evidence Actually Establishes.”

The section included a three-column research table covering:

Retrieval practice.
Audience perception.
Stress responses.
Brain processing.

The answer now looked more scientific than the previous two.

It referred to:

Cognitive psychology.
Neuroscience.
Working memory.
Cortisol.
Attentional control.
Dual coding.
A landmark Psychological Science study.
A 2026 McGill University brain study.

But closer examination showed that the stronger scientific presentation was not matched by stronger citation precision.

Real memory research was applied to a different question

AI Mode said retrieval practice created flexible, stress-resistant recall and used this to support internalising ideas rather than memorising exact wording.

Research does exist showing that retrieval practice can improve later recall and may protect learned material against the effects of acute stress. One cited study compared retrieval practice with repeated studying using memory tasks. (PubMed: Retrieval Practice Protects Memory Against Acute Stress)

However, that study did not directly compare:

Memorised presentation openings.
Outline-based presentation openings.
Nervous first-time speakers.
Audience responses to different delivery methods.

AI Mode took a genuine general finding about memory and applied it to a specific public-speaking disagreement.

That may be a reasonable hypothesis.

It is not the same as direct evidence showing that one presentation method is superior.

A “30%” statistic came from a LinkedIn comment

AI Mode claimed:

“Research shows audiences rate speakers as more credible and retain up to 30% more information when a presentation flows from a practiced structural framework rather than a verbatim script.”

The cited LinkedIn page was a TED Conferences post recommending “precision practice” rather than either memorising or improvising.

The figure did not appear in the TED post itself.

It appeared in a comment below the post:

“Studies show listeners rate presenters more credible and recall up to 30% more…”

The comment did not identify the studies, authors, sample or publication.

AI Mode had effectively converted:

An unsupported numerical claim in a LinkedIn comment

into:

“Research shows…”

We could not identify the underlying study from the material provided.

That does not prove the figure is false.

It means the citation did not provide enough evidence to treat it as established research.

This is an important distinction for responsible content creators:

Unverified is not the same as disproved, but it should not be presented as established fact.

A McGill speech-learning study was moved into presentation advice

AI Mode also cited a 2026 McGill University study.

The research examined the importance of auditory and somatosensory processing in learning and retaining new speech patterns. McGill described it as research into how people learn or regain speech and the role of the brain’s sensory-processing cortex. (McGill University: Research Findings Challenge Assumptions About Speech Learning)

AI Mode used this to suggest that:

“A rigid, unfeeling script fights natural brain processing.”

But the study did not investigate:

Memorised presentations.
Emotional delivery.
Scripts versus outlines.
Public-speaking anxiety.
Audience engagement.

The study was real.

The application was AI Mode’s.

This was a particularly clear example of research-context transplantation: a finding from one scientific setting was used as apparent evidence in another.

The cited coaches still did not agree

AI Mode’s final verdict was:

“Internalise, Don’t Memorise.”

That sounds like a settled middle ground.

But the cited sources used different methods and even different definitions of internalisation.

Throughline Group advises most speakers to internalise the structure and main points rather than memorising every word. (Throughline Group: Should You Memorize Your Presentations?)

Vantage Partners similarly recommends remembering key points and using a framework rather than committing an exact script to memory. (Vantage Partners: Is Memorizing Your Presentation a Bad Idea?)

Effective Presentations recommends memorising the opening and closing while using concept mastery for the body.

By contrast, Moxie Institute argues that properly executed memorisation can improve delivery and recommends complete memorisation in some high-stakes situations. It also advises memorising openers, transitions, stories and closers word for word, even where minimal notes are used. (Moxie Institute: How to Memorize a Speech)

These are meaningfully different positions.

AI Mode acknowledged disagreement but then reorganised it into a clean, evidence-based compromise.

What the three runs revealed

This experiment was too small to prove that every AI Mode answer behaves in the same way.

It was, however, large enough to identify several recurring behaviours within this case study.

1. Recommendation stability with citation instability

Across all three queries, AI Mode maintained a broadly similar preference:

Prepare the opening especially carefully.
Avoid depending on a fully memorised script.
Use a more flexible structure for the main body.

But most of the supporting sources changed between runs.

This suggests that the final recommendation was not simply copied from one permanent source.

It was a stable synthesis that AI Mode could reconstruct from different collections of related material.

2. AI Mode created exact rules from imprecise ranges

The cited pages recommended:

20 seconds.
30–45 seconds.
60 seconds.
60–90 seconds.
The introduction generally.
The opening and closing.
The first one to three minutes.

AI Mode repeatedly converted this range into one clear prescription.

First:

Memorise 60 seconds.

Then:

Memorise 60–90 seconds.

The precision made the answer easier to follow.

It also concealed the variation in the underlying recommendations.

3. Relevant citations were not always precise citations

Many cited pages were genuinely relevant to public speaking, rehearsal or memorisation.

That does not mean they directly supported every nearby sentence.

The citation blocks often behaved more like:

“These pages discuss the surrounding subject.”

They did not always behave like traditional footnotes that proved the exact preceding claim.

4. Cautious language became confident language

The responses repeatedly intensified the source material.

Examples included:

“A prepared opening may help” becoming “vastly superior.”
“Anxiety often peaks early” becoming “adrenaline peaks the moment you start.”
“Forgetting can interrupt recall” becoming “forgetting one word can cause a total blank.”
“A polished opening creates a good impression” becoming “commands the room.”
A general memory finding becoming evidence that one presentation method is scientifically superior.

This is important because certainty changes meaning.

There is a substantial difference between:

“This can help some nervous speakers”

and:

“A nervous beginner should do this.”

5. Advice moved between contexts

The answers reused ideas from:

Slides.
Cue cards.
General study notes.
Memory experiments.
Speech-motor learning.
Executive coaching.
TED-style talks.

AI Mode then combined them into one preparation system for nervous beginners.

That synthesis can be useful.

But it should not be mistaken for a method that has necessarily been tested as a complete package.

6. Genuine disagreement was converted into consensus

The sources included coaches who recommended:

Avoiding word-for-word memorisation.
Memorising only the opening and closing.
Internalising ideas and structure.
Memorising a script deeply enough to sound natural.
Fully memorising some high-stakes presentations.

AI Mode initially concealed much of that tension.

When directly challenged, it acknowledged the different positions but still produced one clean verdict.

This is understandable: users generally want an answer.

But the lost disagreement may contain the most useful part of the subject.

7. A more scientific-looking answer was not necessarily better evidenced

The third answer contained more academic terminology and more references to research than the first two.

It also contained the clearest sourcing problems:

A public-speaking conclusion derived from general memory research.
A “30%” statistic traced to an unsupported LinkedIn comment.
A speech-motor learning study applied to scripted presentation delivery.
A single evidence table that mixed coaching opinion, scientific findings and AI inference.

The answer looked more authoritative because it used the language of science.

That is not the same as having stronger direct evidence.

Does this mean the AI Mode advice was wrong?

Not necessarily.

The hybrid recommendation may be excellent advice for many nervous speakers:

Prepare the opening carefully.
Understand the main ideas.
Use brief notes.
Practise repeatedly.
Avoid becoming dependent on one exact chain of words.

Our experiment did not test people giving presentations, so we cannot determine which method actually produces the best results.

What we can say is:

The citations did not always prove the recommendation with the precision and certainty used in the answer.

That distinction is central to evidence-led publishing.

A responsible writer should separate:

What a source directly found.
What an experienced practitioner recommends.
What appears across several sources.
What the writer has personally tested.
What remains an inference.
What is still uncertain.

Why this matters to hobby website builders

At first glance, this experiment might appear to be about public speaking.

Its real significance is much wider.

AI Mode is designed to search across subtopics and bring material together from multiple pages. Google describes this as “query fan-out”: the system can issue several related searches and combine information from different sources into one response.

This means AI Mode is very good at synthesising the existing web.

If ten pages all repeat similar general advice, AI Mode can often summarise that advice without requiring the user to visit any of them.

A new hobby article that merely repeats the same material may therefore struggle to offer anything beyond the synthesis.

The opportunity lies in the parts AI Mode handles less cleanly.

Opportunity one: Preserve the disagreement

When credible sources disagree, do not rush to hide the tension.

Explain:

Why one expert recommends one approach.
Why another expert disagrees.
Whether they are discussing different situations.
Whether they define the key terms differently.
What evidence exists for each position.
What remains a matter of experience or judgment.

A beginner telescope article could examine why some people recommend a refractor while others recommend a Dobsonian.

The useful contribution is not simply declaring one winner.

It may be explaining that the answer changes according to:

Storage space.
Balcony width.
Light pollution.
Physical lifting ability.
Setup time.
Interest in planets versus deep-sky objects.

AI Mode may compress those positions into “a Dobsonian is generally best.”

A specialist hobby article can preserve the conditions under which that advice changes.

Opportunity two: Test the claim yourself

Our citation audit could identify where the evidence was weak.

A stronger next step would be to conduct a practical test.

For example, someone writing about public speaking could recruit a small group of beginners and compare:

A fully memorised opening.
An outline-only opening.
A rehearsed but non-verbatim opening.

The writer could document:

Preparation time.
Number of pauses.
Number of lost places.
Self-reported anxiety.
Audience ratings.
Speaker preference.
What happened when an interruption was introduced.

That would add something the current collection of coaching articles does not provide.

The same principle applies across hobby subjects.

A telescope site could compare two beginner telescopes from a narrow balcony.

A gardening site could test two watering routines over the same period.

A chess site could compare two study plans and publish the training records.

A craft site could compare materials and document failures as well as successes.

A cleaning article could show the product used, the starting condition, the process and the result.

You do not need a laboratory to create useful evidence.

You do need to show what was actually done.

Opportunity three: Distinguish evidence from judgment

Hobby content often contains a mixture of:

Facts.
Personal preferences.
Rules of thumb.
Manufacturer claims.
Community experience.
Individual tests.
Professional recommendations.

Do not present them all in the same voice.

Instead, write clearly:

“The manufacturer states…”

“Three of the coaches we reviewed recommend…”

“The scientific study found…”

“The study did not test…”

“In my own trial…”

“My interpretation is…”

“This may not apply when…”

That language helps readers understand the strength of each statement.

It also makes the content easier for search systems to interpret accurately.

Opportunity four: Place evidence beside the claim

Avoid placing ten citations after a long section and leaving the reader to work out which source supports which sentence.

Where possible:

Put the link immediately after the claim.
Link to the original research rather than someone mentioning it.
Identify when the source is a practitioner opinion.
Avoid presenting a social-media comment as research.
Do not use one citation to cover several unrelated claims.
Explain when you are combining multiple sources into your own conclusion.

The objective is not to make the page look heavily cited.

The objective is to make every important claim traceable.

Opportunity five: Investigate attractive statistics

Precise percentages create authority.

They also deserve extra scrutiny.

When you encounter a claim such as:

“Audiences remember 30% more…”

Ask:

Which study produced the figure?
Who conducted it?
How many people participated?
What was compared?
Was the result statistically significant?
Was the study about presentations at all?
Is the source reporting the study or merely repeating the number?

If the original evidence cannot be found, say so.

That investigation can itself become valuable content.

Opportunity six: Explain where a study stops

One of the clearest failures in our experiment was not the use of false research.

It was the extension of real research beyond what it tested.

A good hobby writer should explain the boundary.

For example:

“This study found that retrieval practice protected recall of learned material under acute stress. It did not compare memorised and outline-based presentation openings, so applying it to public speaking remains an inference.”

That sentence adds more value than simply attaching the study to a recommendation.

It tells the reader both what we know and what we do not know.

Opportunity seven: Become the source that documents the missing step

AI Mode can combine existing pages.

It cannot retrieve evidence that nobody has created.

This is the strongest opportunity for hobby website builders.

Do the test.

Take the photographs.

Record the measurements.

Save the screenshots.

Document the failure.

Show the setup.

Explain the conditions.

Publish the raw result.

Separate what happened from what you think it means.

A small hobby website may not have the authority or publishing volume of a major media brand.

But it can still become the clearest source for one tightly defined practical question.

A practical evidence-led content method

This experiment suggests a repeatable approach for building hobby content.

Step 1: Research what already exists

Read the current articles, forums, videos, manuals and AI Mode answers.

Do not begin by rewriting the first few pages you find.

Map the existing positions.

Step 2: Identify the tension

Look for:

Contradictory advice.
Unsupported rules.
Repeated statistics with no original source.
Different recommendations for different users.
Claims that appear more certain than the evidence.
Advice transferred from one context to another.

This is often where the new article begins.

Step 3: Decide what can be tested

Create a practical comparison, observation or documented process.

It can be small.

The important point is that you actually do it.

Step 4: Preserve the evidence

Keep:

Before-and-after photographs.
Screenshots.
Dates.
Measurements.
Products and equipment used.
Environmental conditions.
Failed attempts.
Unexpected results.
Notes made during the process.

Do not try to reconstruct the evidence after writing the article.

Step 5: Separate the layers

Organise the final post into:

What existing sources say.
Where they disagree.
What you tested.
What happened.
What you infer from the result.
What the test cannot establish.
What the reader should consider in their own situation.

Step 6: Cite precisely

Attach the strongest available source to the exact claim it supports.

Do not manufacture the appearance of authority through citation volume.

Step 7: Present it clearly

The crawler matters, but the human reader comes first.

Use:

Descriptive headings.
Short sections.
Clear photographs.
Useful captions.
Simple comparison boxes.
Honest limitations.
Direct conclusions.

Optimise the communication.

Do not manufacture the signals.

This is not a special AI Mode optimisation trick

Google’s own guidance says there are no additional technical requirements or special optimisations needed to appear in AI Overviews or AI Mode beyond the established requirements for Google Search.

Google continues to recommend helpful, reliable, people-first content and says AI features may surface a broader and more diverse set of supporting links.

The lesson from this experiment is therefore not:

Add more citations so AI Mode will trust the page.

Nor is it:

Use scientific language to make the article appear authoritative.

The lesson is:

Create content that deserves to be used as evidence.

That means contributing something the existing synthesis cannot easily supply:

A documented test.
A clearly explained disagreement.
An original comparison.
A source audit.
A real failure.
A set of measurements.
A situation-specific judgment.
Evidence that the writer actually did the work.

The limits of our experiment

This was a focused case study involving:

One subject.
Three closely related queries.
AI Mode responses produced during one testing period.
A manual review of the cited pages.

It does not prove that every AI Mode answer handles citations in the same way.

AI-generated responses can change according to:

Query wording.
Location.
Personalisation.
Product updates.
Source-page changes.
Search-index changes.
Follow-up context.

Some of the cited pages may also be edited after the experiment.

Our conclusions should therefore remain proportionate.

We observed repeated behaviours in these three responses.

We did not reverse-engineer Google’s entire system.

Final conclusion

The AI Mode answers in this experiment were not random collections of invented advice.

They found relevant pages, recognised recurring ideas and produced a practical recommendation.

But the citations did not always function as readers might expect.

Across the three runs, AI Mode:

Maintained a broadly stable recommendation while changing most of its sources.
Converted a range of expert advice into precise rules.
Strengthened cautious wording.
Attached broad citation groups to specific claims.
Moved advice between related contexts.
Used genuine research outside the setting it directly tested.
Turned an unsupported LinkedIn comment into a research statistic.
Acknowledged disagreement before constructing a cleaner consensus.

The most important conclusion is therefore not that AI Mode gave useless advice.

It is this:

A coherent, sensible and heavily cited AI answer can still contain conclusions that the cited sources do not directly establish.

For hobby website builders, that creates a clear direction.

Do not attempt to compete with AI Mode by producing another generic summary of the same pages it can already read.

Research what exists.

Find the disagreement or evidential gap.

Do the work.

Preserve the evidence.

Distinguish observation from inference.

Cite the exact claim.

Present everything as clearly as possible.

AI Mode can synthesise the existing web.

Your opportunity is to create the evidence that deserves to become part of that synthesis.