Google AI Mode can produce a polished answer supported by an impressive-looking list of citations.
But how stable is that answer?
Does asking the same question repeatedly produce the same recommendation? Do the cited pages support the precise claims being made? And what happens when the question is reworded to emphasise a different part of the problem?
I designed a small experiment to investigate those questions.
The subject was stage fright. More specifically, I wanted to know whether a nervous beginner should memorise the opening of a presentation, use bullet-point notes or avoid memorising exact wording.
This was a useful test because there is no single mechanical answer. Unlike cleaning a microwave or changing a tyre, public speaking involves individual psychology, personal preference and competing schools of thought.
Some speaking coaches recommend memorising the opening.
Others warn that word-for-word memorisation can make anxiety worse.
That tension gave AI Mode something real to resolve.
What I tested
I began with this query:
I get severe stage fright before giving a presentation. Should I memorise the first 60 seconds, use bullet-point notes, or avoid memorising anything? Explain the advantages and risks of each approach and recommend the best option for a nervous beginner.
I entered that exact query five times, each time in a fresh AI Mode conversation.
I then tested three variations:
Is memorising the opening minute of a presentation helpful for someone with stage fright, or can it make the anxiety worse?
For a very nervous public speaker, is it better to learn the introduction word for word or speak from brief notes?
What should a nervous beginner prepare before a presentation: a memorised opening, an outline, or neither?
For every run, I copied:
- the complete AI Mode response;
- its numbered citation markers;
- and the full list of linked sources.
Across the eight responses, AI Mode displayed 155 citation entries. Some sources were repeated, so this was not 155 unique pages.
I then compared the recommendations and inspected important underlying sources to see whether they directly supported the claims beside which they appeared.
This was a small observational experiment, not an academic study. It used one subject, one account and searches conducted within a short period. Its purpose was to observe AI Mode’s behaviour and identify useful content-development lessons.
The headline results
The broad answer was highly consistent.
Every response recommended some combination of:
- preparing the opening particularly carefully;
- creating a clear structure;
- rehearsing aloud;
- and using brief prompts or bullet-point notes.
AI Mode never recommended that a very nervous beginner should simply stand up and improvise without preparation.
However, its advice about exact memorisation was not stable.
| Run | Headline recommendation |
|---|---|
| 1 | Memorise the first 60 seconds, then use bullet points |
| 2 | Memorise the first 60 seconds, then use bullet points |
| 3 | Memorise the first 60 seconds, then use bullet points |
| 4 | Memorise the first 60 seconds, then use bullet points |
| 5 | Memorise the opening concept, not the exact words |
| Variation 1 | Internalise the opening; do not script it verbatim |
| Variation 2 | Memorise the introduction word for word, then use notes |
| Variation 3 | Make an outline the priority and memorise two opening sentences |
The broad compromise remained stable.
The precise rule did not.
That distinction matters.
It would be inaccurate to say that AI Mode produced eight completely different answers. It repeatedly found the same general area of agreement.
But it would also be inaccurate to say it gave the same answer every time. The practical difference between memorising 60 seconds perfectly and deliberately avoiding exact wording could be very important to someone who is already afraid of freezing.
Runs 1 to 4: a confident rule emerges
The first four identical runs strongly recommended word-for-word memorisation of the opening.
Run 1 began:
Memorising the first 60 seconds is the best approach for a nervous beginner. It provides a reliable bridge over the initial spike of adrenaline when stage fright is at its worst. [1]
It later recommended:
Script the first 60 seconds: Memorise your opening lines until they are pure muscle memory.
Run 2 was similarly decisive:
Yes, memorising the first 60 seconds is the best strategy for a nervous beginner.
Run 3 added a more scientific-sounding explanation:
By memorising just the first minute, you create an automatic “autopilot” mode. This carries you through the highest spike of anxiety.
Run 4 instructed the reader to:
Memorise the first 60 seconds perfectly.
Four consecutive runs therefore transformed a broad preparation principle into a very precise formula:
Memorise exactly one minute, word for word, and then switch to bullet points.
Some of the cited material supported this approach.
For example, an Oxford University Press teaching article advises speakers to know the first 60 seconds extremely well because the opening is commonly the most nerve-racking part. It says that after a smooth start, anxiety often begins to dissipate. However, it does not explicitly say that every word must be memorised, and it also advises people to experiment with different strategies to discover what works for them. Read the Oxford article.
This is an important distinction.
The source says, in effect:
Become extremely familiar with the opening.
AI Mode converted that into:
Memorise the first 60 seconds perfectly.
The two ideas are related, but they are not identical.
Run 5: the answer changes without the question changing
The fifth identical run broke the pattern.
Its final recommendation was:
Memorise your opening concept, not the exact words. Know exactly what your first three sentences will achieve, but don’t script them word-for-word to avoid freeze-frames.
This was the first response to make the central distinction explicit:
Thorough preparation is not necessarily the same as exact memorisation.
The overall hybrid remained. AI Mode still recommended a carefully prepared opening followed by structured bullet points.
But the most important implementation detail had reversed.
In the first four runs, exact wording was presented as the solution.
In the fifth, exact wording was presented as part of the risk.
This showed that even without changing the query, the synthesis was not completely fixed.
The sources contained a genuine disagreement
The variation was not necessarily random. The cited web contained real disagreement.
The National Social Anxiety Center advises speakers to practise but not memorise. It warns that exact memorisation can create a false sense of security: forgetting one phrase may increase anxiety and make a mental blank more likely. It recommends internalising the flow and using bullet points instead. Read the National Social Anxiety Center article.
Vantage Partners gives similar advice. It argues that a speaker who concentrates on reproducing a script may sound robotic, become less adaptable and struggle to recover after forgetting one word. It recommends using a framework rather than depending on one exact sequence of sentences. Read the Vantage Partners article.
Other practitioners take a more conditional view.
Professional speaker Shola Kaye explains that exact memorisation may initially give a beginner confidence. She suggests that the introduction, conclusion and difficult facts may deserve extra memorisation, while also warning that learning an entire presentation word for word can damage audience connection and make interruptions difficult to recover from. Read Shola Kaye’s article.
These are not three versions of precisely the same recommendation.
They represent different positions:
- Memorise the opening.
- Rehearse the opening extensively but avoid exact wording.
- Memorise selected sections when useful, but do not rely on a full script.
AI Mode had to decide how to combine them.
Most of the time, it chose a neat hybrid:
Memorised opening plus bullet-point middle.
That may be sensible advice. But presenting it as the best method hides the fact that the underlying sources disagree about what the opening preparation should involve.
Rewording the query changed which side AI Mode favoured
The three query variations produced the strongest evidence that framing affects the synthesis.
Variation 1: asking whether memorisation can make anxiety worse
The query was:
Is memorising the opening minute of a presentation helpful for someone with stage fright, or can it make the anxiety worse?
AI Mode answered:
Memorising the opening minute of a presentation can be highly beneficial for managing stage fright, but it can also backfire and worsen anxiety if done incorrectly.
Its conclusion was:
The Best Approach: Internalize, Don’t Script Verbatim
This was noticeably more cautious than the first four runs.
The question had explicitly invited AI Mode to examine the harm caused by memorisation. The resulting answer gave much more weight to sources warning about perfectionism, mental blanks and robotic delivery.
Variation 2: directly contrasting word-for-word learning with notes
The query was:
For a very nervous public speaker, is it better to learn the introduction word for word or speak from brief notes?
This time AI Mode swung back:
For a very nervous public speaker, it is highly recommended to memorise your introduction word-for-word, but switch to brief bulleted notes for the remainder of the speech.
The same broad evidence landscape was available, but the headline recommendation once again favoured exact memorisation.
Variation 3: making the outline a named option
The final query was:
What should a nervous beginner prepare before a presentation: a memorised opening, an outline, or neither?
AI Mode now answered:
A nervous beginner should prepare an outline.
It reduced memorisation to a smaller supporting technique:
Memorise just the first two sentences to build momentum.
This was the clearest change in hierarchy.
In earlier runs, the memorised minute was the central safety mechanism and notes supported the rest of the talk.
Here, the outline became the central safety mechanism and two prepared sentences supported the start.
The For Dummies article cited in this run closely supports the outline-first interpretation. It recommends rehearsing the opening more than other sections, keeping short memory-jogging notes and internalising the message rather than memorising a script. Read the article.
Presentation coach Michelle Bowden similarly distinguishes rehearsal from rote learning. She recommends saying the message differently during each rehearsal, while still spending extra time on the opening and close. Read Michelle Bowden’s article.
The final variation therefore surfaced a coherent position that had existed within the source material throughout:
Rehearse the beginning heavily, know the structure and use prompts, but do not become dependent on one exact verbal sequence.
What this suggests about how AI Mode operates
We cannot see AI Mode’s private reasoning or know precisely why it selected one source or conclusion over another.
We can, however, compare the experiment with Google’s own explanation of the system.
Google says AI Mode can divide a question into subtopics and search for them simultaneously. Its documentation describes a “query fan-out” process through which related searches are issued across different subtopics and data sources. Google also says that responses and supporting links can vary because different models and techniques may be used.
That fits what appeared in this experiment.
The recommendation was often similar, but the source lists changed considerably. AI Mode appeared to gather pages dealing with several connected issues:
- the initial anxiety spike;
- memorising introductions;
- the dangers of rote recall;
- bullet-point notes;
- eye contact;
- mental blanks;
- rehearsal;
- and audience connection.
It then assembled those ideas into one readable answer.
The important observation is that AI Mode was not simply retrieving one established verdict.
It was constructing a verdict from a changing collection of material.
The framing of the question appeared to influence which part of that material became the main conclusion.
When asked for the “best option,” AI Mode usually gave a decisive rule.
When asked whether memorisation might be harmful, it elevated the warnings.
When an outline was offered as an explicit alternative, the outline became the main recommendation.
Google itself advises users to check important information in more than one place and to ask multiple versions of a question, including requests for different opinions. This experiment demonstrates why that advice is useful.
A citation is not the same as proof of every nearby claim
The experiment also showed why citation checking matters.
A response containing 20 or 30 links can look thoroughly evidenced. But several different relationships can exist between an AI claim and the citation beside it.
A source may:
- directly support the exact claim;
- support only the broad theme;
- support one bullet in a cluster but not the others;
- recommend something slightly different;
- contradict part of the conclusion;
- or be inaccessible and therefore difficult to verify.
For example, the Oxford source supports becoming exceptionally familiar with the beginning. It does not clearly establish that memorising exactly 60 seconds word for word is universally best.
The National Social Anxiety Center directly advises against exact memorisation.
Shola Kaye allows for selective memorisation while emphasising its risks.
Vantage Partners recommends frameworks instead of scripts.
AI Mode was therefore entitled to identify a useful area of agreement:
Prepare the opening more carefully than the rest, know the flow and use prompts.
But the precise rules it added—60 seconds, two sentences, three sentences or three-to-four-word triggers—were much less stable than the polished answers made them appear.
Citation quantity did not automatically equal citation precision.
AI Mode was strongest at finding the broad consensus
The experiment should not be read as evidence that AI Mode failed.
Its broad conclusion was sensible and remarkably consistent:
- do not arrive unprepared;
- devote extra rehearsal to the opening;
- organise the presentation clearly;
- use brief prompts;
- practise aloud;
- and avoid trying to improvise the entire talk while severely anxious.
That is useful synthesis.
The weakness appeared when AI Mode converted a broad consensus into an exact universal formula.
The web sources agreed that the opening deserves special preparation.
They did not agree that exactly 60 seconds should always be memorised word for word.
AI Mode repeatedly crossed the gap between those two statements without making the uncertainty fully visible.
The content opportunity is inside the disagreement
This is where the experiment becomes useful for hobby website builders.
A generic article called Ten Ways to Overcome Stage Fright would probably repeat information that AI Mode can already summarise very effectively.
It might mention breathing, practice, eye contact, positive thinking and bullet-point notes.
That material is easy to synthesise because it already exists across many websites.
A stronger content opportunity is:
Should a nervous speaker memorise the first 60 seconds or avoid word-for-word learning?
That question contains a real unresolved tension.
A useful human-written article could:
- explain why some coaches favour memorisation;
- explain why others warn against it;
- separate memorising ideas from memorising sentences;
- compare different preparation methods;
- interview speakers or coaches;
- run a practical test;
- and help different types of nervous speakers choose an approach.
The opportunity is not merely to provide more words.
It is to uncover the distinction that the broad synthesis tends to flatten.
What hobby website builders should do differently
1. Research the answer AI can already provide
Before writing, ask AI Mode several versions of the question.
Do not do this to copy the response.
Do it to understand:
- the broad consensus;
- the common sources;
- the standard advice;
- the repeated phrases;
- and what AI Mode can already explain without your help.
That gives you a map of the existing information landscape.
2. Look for tensions rather than missing keywords
The strongest opportunity may not be a topic that nobody has mentioned.
It may be a disagreement that existing summaries smooth over.
Examples might include:
- Is a reflector or refractor genuinely easier for a first telescope?
- Should sourdough dough be handled on a timed schedule or judged by visible fermentation?
- Is a pond pump necessary for a small wildlife pond?
- Should vegetable seedlings be watered lightly every day or deeply less often?
- Is an expensive specialist tool genuinely better for a beginner than a simple alternative?
These are useful because different people may have reached different conclusions under different conditions.
The writer’s task is to explain why.
3. Inspect the citations instead of trusting the citation count
Click the links.
For each important claim, ask:
- Does this page directly support the wording?
- Is it reporting a test or giving one person’s opinion?
- Is the source discussing the exact situation?
- Has AI Mode made the advice more precise than the source?
- Are different kinds of evidence being treated as equivalent?
- Does another cited source disagree?
The citation audit may itself reveal the article.
4. Do something that creates new evidence
Research can uncover the tension.
First-hand work can help resolve it.
A telescope site could compare how long two beginner mounts take to set up and point at the Moon.
A gardening site could test two watering schedules using the same plants and compost.
A baking site could photograph dough at different stages and compare the finished loaves.
A cleaning site could try vinegar, lemon and plain water on similar microwave stains and record which method actually makes a difference.
The writer is no longer merely choosing between existing opinions.
The writer is adding another piece of inspectable evidence.
5. Preserve the research record
Keep:
- the exact query;
- the date;
- the complete AI Mode response;
- the original citation list;
- screenshots;
- your notes from checking the sources;
- photographs from any practical test;
- measurements and results;
- and evidence of failures or unexpected outcomes.
This allows the reader to see how the conclusion was reached.
The evidence should not be hidden behind the final answer.
It should be part of the value of the page.
6. Cover the underlying decision, not every wording variation
The experiment showed that query wording matters.
That does not mean a website should publish separate near-duplicate posts for every possible wording.
A better article would address the complete underlying decision:
Memorising a presentation opening: when it helps, when it hurts and what nervous beginners should do instead.
Google’s current guidance specifically warns against creating separate content for every possible query variation primarily to influence search or generative AI results. It instead recommends unique, useful, non-commodity content organised for human readers.
7. Be the source that preserves the nuance
AI synthesis is designed to make complicated information easier to consume.
That usually requires compression.
Compression can remove:
- uncertainty;
- conditions;
- exceptions;
- source disagreement;
- individual circumstances;
- and the difference between a broad principle and an exact rule.
A strong hobby article can preserve those details without becoming confusing.
The goal is not to make the answer unnecessarily complicated.
It is to prevent an important distinction from disappearing.
This fits the evidence-led content method
The experiment reinforces a straightforward content process:
Research what already exists, contribute something genuinely new, do the work, preserve the evidence and present it as clearly as possible.
The copied AI Mode response is not the article.
It is part of the research record.
The citation list is not proof that the synthesis is correct.
It is an invitation to inspect the underlying material.
The writer’s original contribution comes from:
- comparing the outputs;
- checking the citations;
- identifying the hidden disagreement;
- testing the advice where possible;
- and presenting a conclusion that accurately reflects the evidence.
Google’s own current guidance encourages creators to provide original research and analysis, demonstrate first-hand experience and avoid simply summarising what others have already said. It also says there is no special technical trick required to appear in AI Mode beyond the established foundations of useful, accessible, people-first content.
Final conclusion
This experiment suggests that AI Mode is good at finding a practical middle ground across many sources.
It identified a stable broad principle:
A nervous beginner should prepare the opening carefully, rehearse the structure and use brief prompts for support.
But it was less stable when turning that principle into a precise instruction.
Depending on the run and the wording of the query, the reader was told to:
- memorise 60 seconds perfectly;
- memorise the opening word for word;
- memorise only the concept;
- internalise rather than script;
- memorise two sentences;
- or make an outline the primary method.
The source material contained reasons for each variation.
AI Mode’s role was to decide which reasons to emphasise and how to combine them.
That is the central finding:
AI Mode does not merely uncover one fixed answer. It constructs a useful synthesis, and the framing of the question can influence how competing evidence is resolved.
For hobby website builders, that should not be discouraging.
It points towards the work that still matters.
Do not try to beat AI Mode by producing another broad summary of information it can already combine.
Find the tension.
Inspect the sources.
Do the experiment.
Preserve the evidence.
Then create the page that explains what the polished synthesis left unresolved.